2023Activity reportProject-TeamMAGNET

7.1.1 CoRTeX

• Name:
  Python library for noun phrase COreference Resolution in natural language TEXts
• Keyword:
  Natural language processing
• Functional Description:
  CoRTex is a LGPL-licensed Python library for Noun Phrase coreference resolution in natural language texts. This library contains implementations of various state-of-the-art coreference resolution algorithms, including those developed in our research. In addition, it provides a set of APIs and utilities for text pre-processing, reading the CONLL2012 and CONLLU annotation formats, and performing evaluation, notably based on the main evaluation metrics (MUC, B-CUBED, and CEAF). As such, CoRTex provides benchmarks for researchers working on coreference resolution, but it is also of interest for developers who want to integrate a coreference resolution within a larger platform. It currently supports use of the English or French language.
• Contact:
  Pascal Denis
• Participant:
  Pascal Denis
• Partner:
  Orange Labs

7.1.2 Mangoes

• Name:
  MAgnet liNGuistic wOrd vEctorS
• Keywords:
  Word embeddings, NLP
• Functional Description:

  Mangoes is a toolbox for constructing and evaluating static and contextual token vector representations (aka embeddings). The main functionalities are:

  - Contextual embeddings: Access a large collection of pretrained transformer-based language models, Pre-train a BERT language model on a corpus, Fine-tune a BERT language model for a number of extrinsic tasks, Extract features/predictions from pretrained language models.

  - Static embeddings: Process textual data and compute vocabularies and co-occurrence matrices. Input data should be raw text or annotated text, Compute static word embeddings with different state-of-the art unsupervised methods, Propose statistical and intrinsic evaluation methods, as well as some visualization tools, Generate context dependent embeddings from a pretrained language model.

  Future releases will include methods for injecting lexical and semantic knowledge into token and multi-model embeddings, and interfaces into common external knowledge resources.

• URL:
  https://­gitlab.­inria.­fr/­magnet/­mangoes
• Contact:
  Nathalie Vauquier

7.1.3 metric-learn

• Keywords:
  Machine learning, Python, Metric learning
• Functional Description:

  Distance metrics are widely used in the machine learning literature. Traditionally, practicioners would choose a standard distance metric (Euclidean, City-Block, Cosine, etc.) using a priori knowledge of the domain. Distance metric learning (or simply, metric learning) is the sub-field of machine learning dedicated to automatically constructing optimal distance metrics.

  This package contains efficient Python implementations of several popular metric learning algorithms.

• URL:
  https://­github.­com/­scikit-learn-contrib/­metric-learn
• Contact:
  Aurelien Bellet
• Partner:
  Parietal

7.1.4 MyLocalInfo

• Keywords:
  Privacy, Machine learning, Statistics
• Functional Description:
  Decentralized algorithms for machine learning and inference tasks which (1) perform as much computation as possible locally and (2) ensure privacy and security by avoiding that personal data leaves devices.
• Contact:
  Nathalie Vauquier

7.1.5 declearn

• Keyword:
  Federated learning
• Scientific Description:

  declearn is a python package providing with a framework to perform federated learning, i.e. to train machine learning models by distributing computations across a set of data owners that, consequently, only have to share aggregated information (rather than individual data samples) with an orchestrating server (and, by extension, with each other).

  The aim of declearn is to provide both real-world end-users and algorithm researchers with a modular and extensible framework that:

  - builds on abstractions general enough to write backbone algorithmic code agnostic to the actual computation framework, statistical model details or network communications setup - designs modular and combinable objects, so that algorithmic features, and more generally any specific implementation of a component (the model, network protocol, client or server optimizer...) may easily be plugged into the main federated learning process - enabling users to experiment with configurations that intersect unitary features - provides with functioning tools that may be used out-of-the-box to set up federated learning tasks using some popular computation frameworks (scikit- learn, tensorflow, pytorch...) and federated learning algorithms (FedAvg, Scaffold, FedYogi...) - provides with tools that enable extending the support of existing tools and APIs to custom functions and classes without having to hack into the source code, merely adding new features (tensor libraries, model classes, optimization plug-ins, orchestration algorithms, communication protocols...) to the party.

  Parts of the declearn code (Optimizers,...) are included in the FedBioMed software.

  At the moment, declearn has been focused on so-called "centralized" federated learning that implies a central server orchestrating computations, but it might become more oriented towards decentralized processes in the future, that remove the use of a central agent.

• Functional Description:
  This library provides the two main components to perform federated learning: - the client, to be run by each participant, performs the learning on local data et releases only the result of the computation - the server orchestrates the process and aggregates the local models in a global model
• URL:
  https://­gitlab.­inria.­fr/­magnet/­declearn/­declearn2
• Contact:
  Aurelien Bellet
• Participants:
  Paul Andrey, Aurelien Bellet, Nathan Bigaud, Marc Tommasi, Nathalie Vauquier
• Partner:
  CHRU Lille

7.1.6 fairgrad

• Name:
  FairGrad: Fairness Aware Gradient Descent
• Keywords:
  Fairness, Fair and ethical machine learning, Machine learning, Classification
• Functional Description:
  FairGrad is an easy to use general purpose approach in Machine Learning to enforce fairness in gradient descent based methods
• URL:
  https://­github.­com/­saist1993/­fairgrad
• Authors:
  Gaurav Maheshwari, Michael Perrot
• Contact:
  Michael Perrot

7.1.7 tasksource

• Name:
  tasksource
• Keyword:
  Natural language processing
• Functional Description:
  tasksource streamlines interchangeable datasets usage to scale evaluation or multi-task learning. All implemented preprocessings are in tasks.py or tasks.md. A preprocessing is a function that accepts a dataset and returns the standardized dataset. Preprocessing code is concise and human-readable.
• URL:
  https://­github.­com/­sileod/­tasksource
• Publication:
  hal-04099649v1
• Contact:
  Damien Sileo

7.1.8 Voice Transformer 2

• Keywords:
  Speech, Privacy
• Scientific Description:

  The implemented method is inspired from the speaker anonymisation method proposed in [Fan+19], which performs voice conversion based on x-vectors [Sny+18], a fixed-length representation of speech signals that form the basis of state-of-the-art speaker verification systems. We have brought several improvements to this method such as pitch transformation, and new design choices for x-vector selection

  [Fan+19] F. Fang, X. Wang, J. Yamagishi, I. Echizen, M. Todisco, N. Evans, and J.F. Bonastre. “Speaker Anonymization Using x-vector and Neural Waveform Models”. In: Proceedings of the 10th ISCA Speech Synthesis Workshop. 2019, pp. 155–160. [Sny+18] D. Snyder, D. Garcia-Romero, G. Sell, D. Povey, and S. Khudanpur. “X-vectors: Robust DNN embeddings for speaker recognition”. In: Proceedings of ICASSP 2018 - 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). 2018, pp. 5329–5333.

• Functional Description:
  Voice Transformer increases the privacy of users of voice interfaces by converting their voice into another person’s voice without modifying the spoken message. It ensures that any information extracted from the transformed voice can hardly be traced back to the original speaker, as validated through state-of-the-art biometric protocols, and it preserves the phonetic information required for human labelling and training of speech-to-text models.
• News of the Year:
  A transfer contract was signed with the startup Nijta.
• Contact:
  Nathalie Vauquier
• Participants:
  Brij Mohan Lal Srivastava, Nathalie Vauquier, Emmanuel Vincent, Marc Tommasi

Data Provenance Initiative

Damien Sileo co-founded the Data Provenance Initiative (DPI), a project focused on transparency about the natural language processing datasets used for the final stages of large language models training. This stage, known as fine-tuning, involves converting a multitude of existing datasets into input-output pairs for language modeling. The DPI addresses a key oversight in this process: the neglect of metadata during dataset aggregation. One notable achievement of DPI is leveraging 37 to recover missing licenses and attributions, revealing inaccuracies in licenses on prominent platforms like GitHub and HuggingFace (the most popular hub for NLP datasets). DPI also traced the origins of texts and annotated datasets with additional details like date, task types, and geographical origins, facilitating comprehensive audits of datasets in natural language processing.

• Contact person: Damien Sileo
• Datasets: https://www.dataprovenance.org/ (and the associated github link available at this address)

Tasksource

Commitment to public dataset release is essential for open and reproducible science. However, datasets often require preprocessing for effective model training or evaluation. The Tasksource project 40 enhances dataset usability by offering one-line commands for loading datasets into standardized formats, ensuring smooth training and evaluation. Tasksource introduces a system for annotating preprocessing steps, provides numerous preprocessing annotations, and has uploaded over a hundred datasets to HuggingFace/tasksource, sourced from GitHub, academic websites, or directly from authors.

• Contact person: Damien Sileo
• Datasets: https://github.com/sileod/tasksource

Exploring Category Structure with Contextual Language Models and Lexical Semantic Networks 28

Recent work on predicting category structure with distributional models, using either static word embeddings (Heyman and Heyman, 2019) or contextualized language models (CLMs) (Misra et al., 2021), report low correlations with human ratings, thus calling into question their plausibility as models of human semantic memory. In this work, we revisit this question testing a wider array of methods for probing CLMs for predicting typicality scores. Our experiments, using BERT (Devlin et al., 2018), show the importance of using the right type of CLM probes, as our best BERT-based typicality prediction methods substantially improve over previous works. Second, our results highlight the importance of polysemy in this task: our best results are obtained when using a disambiguation mechanism. Finally, additional experiments reveal that Information Contentbased WordNet (Miller, 1995), also endowed with disambiguation, match the performance of the best BERT-based method, and in fact capture complementary information, which can be combined with BERT to achieve enhanced typicality predictions.

A Tale of Two Laws of Semantic Change: Predicting Synonym Changes with Distributional Semantic Models 23

Lexical Semantic Change is the study of how the meaning of words evolves through time. Another related question is whether and how lexical relations over pairs of words, such as synonymy, change over time. There are currently two competing, apparently opposite hypotheses in the historical linguistic literature regarding how synonymous words evolve: the Law of Differentiation (LD) argues that synonyms tend to take on different meanings over time, whereas the Law of Parallel Change (LPC) claims that synonyms tend to undergo the same semantic change and therefore remain synonyms. So far, there has been little research using distributional models to assess to what extent these laws apply on historical corpora. In this work, we take a first step toward detecting whether LD or LPC operates for given word pairs. After recasting the problem into a more tractable task, we combine two linguistic resources to propose the first complete evaluation framework on this problem and provide empirical evidence in favor of a dominance of LD. We then propose various computational approaches to the problem using Distributional Semantic Models and grounded in recent literature on Lexical Semantic Change detection. Our best approaches achieve a balanced accuracy above 0.6 on our dataset. We discuss challenges still faced by these approaches, such as polysemy or the potential confusion between synonymy and hypernymy.

WordNet Is All You Need: A Surprisingly Effective Unsupervised Method for Graded Lexical Entailment 29

We propose a simple unsupervised approach which exclusively relies on WordNet (Miller, 1995) for predicting graded lexical entailment (GLE) in English. Inspired by the seminal work of Resnik (1995), our method models GLE as the sum of two information-theoretic scores: a symmetric semantic similarity score and an asymmetric specificity loss score, both exploiting the hierarchical synset structure of Word-Net. Our approach also includes a simple disambiguation mechanism to handle polysemy in a given word pair. Despite its simplicity, our method achieves performance above the state of the art (Spearman $\rho$ = 0.75) on HyperLex (Vulic et al., 2017), the largest GLE dataset, outperforming all previous methods, including specialized word embeddings approaches that use WordNet as weak supervision.

Find-2-Find: Multitask Learning for Anaphora Resolution and Object Localization 27

In multimodal understanding tasks, visual and linguistic ambiguities can arise. Visual ambiguity can occur when visual objects require a model to ground a referring expression in a video without strong supervision, while linguistic ambiguity can occur from changes in entities in action flows. As an example from the cooking domain, ”oil” mixed with ”salt” and ”pepper” could later be referred to as a ”mixture”. Without a clear visual-linguistic alignment, we cannot know which among several objects shown is referred to by the language expression ”mixture”, and without resolved antecedents, we cannot pinpoint what the mixture is. We define this chicken-and-egg problem as visual-linguistic ambiguity. In this paper, we present Find2Find, a joint anaphora resolution and object localization dataset targeting the problem of visual-linguistic ambiguity, consisting of 500 anaphora-annotated recipes with corresponding videos. We present experimental results of a novel end-to-end joint multitask learning framework for Find2Find that fuses visual and textual information and shows improvements both for anaphora resolution and object localization as compared to a strong single-task baseline.

Probing neural language models for understanding of words of estimative probability 31

Words of Estimative Probability (WEP) are phrases used to express the plausibility of a statement. Examples include terms like probably, maybe, likely, doubt, unlikely, and impossible. Surveys have shown that human evaluators tend to agree when assigning numerical probability levels to these WEPs. For instance, the term highly likely equates to a median probability of 0.90$\pm$0.08 according to a survey by Fagen-Ulmschneider (2015). In this study, our focus is to gauge the competency of neural language processing models in accurately capturing the consensual probability level associated with each WEP. Our first approach is utilizing the UNLI dataset (Chen et al., 2020), which links premises and hypotheses with their perceived joint probability p. From this, we craft prompts in the form: ”[PREMISE]. [WEP], [HYPOTHESIS].” This allows us to evaluate whether language models can predict if the consensual probability level of a WEP aligns closely with p. In our second approach, we develop a dataset based on WEP-focused probabilistic reasoning to assess if language models can logically process WEP compositions. For example, given the prompt ”[EVENTA] is likely. [EVENTB] is impossible.”, a wellfunctioning language model should not conclude that [EVENTA&B] is likely. Through our study, we observe that both tasks present challenges to out-of-the-box English language models. However, we also demonstrate that fine-tuning these models can lead to significant and transferable improvements.

MindGames: Targeting Theory of Mind in Large Language Models with Dynamic Epistemic Modal Logic 39

Theory of Mind (ToM) is a critical component of intelligence, yet accurately measuring it continues to be a subject of debate. Prior research has attempted to apply human ToM assessments to natural language processing models using either human-created standardized tests or rule-based templates. However, these methods primarily focus on simplistic reasoning and require further validation. In this study, we utilize dynamic epistemic logic, which has established overlaps with ToM, to generate more intricate problems. We also introduce novel verbalization techniques to express these problems using natural language. Our findings indicate that certain language model scaling (from 70M to 6B and 350M to 174B) does not consistently yield results better than random chance. While GPT-4 demonstrates improved epistemic reasoning capabilities, there is still room for enhancement. Our code and datasets are publicly available here and there.

tasksource: A Dataset Harmonization Framework for Streamlined NLP Multi-Task Learning and Evaluation 40

The HuggingFace Datasets Hub hosts thousands of datasets, offering exciting opportunities for language model training and evaluation. However, datasets for a specific task type often have different schemas, making harmonization challenging. Multi-task training or evaluation necessitates manual work to fit data into task templates. Several initiatives independently tackle this issue by releasing harmonized datasets or providing harmonization codes to preprocess datasets into a consistent format. We identify patterns across previous preprocessing efforts, such as column name mapping and extracting specific sub-fields from structured data in a column. We then propose a structured annotation framework that ensures our annotations are fully exposed and not hidden within unstructured code. We release a dataset annotation framework and dataset annotations for more than 500 English tasks1. These annotations include metadata, such as the names of columns to be used as input or labels for all datasets, which can save time for future dataset preprocessing, regardless of whether our framework is utilized. We fine-tune a multi-task text encoder on all tasksource tasks, outperforming every publicly available text encoder of comparable size in an external evaluation.

Models of Modals 32

Modal verbs in English communicate delicate shades of meaning, there being a large range of verbs both on the necessity side (must, have to, should, ought to, need, need to) and the possibility side (can, may, could, might, be able to). They therefore constitute excellent test ground to apply and compare different methodologies that can lay bare the factors that drive the speaker’s choice of modal verb. This book is not merely concerned with a purely grammatical description of the use of modal verbs, but aims at advancing our understanding of lexical and grammatical units in general and of linguistic methodologies to explore these. It thus involves a genuine effort to compare, assess and combine a variety of approaches. It complements the leading descriptive qualitative work on modal verbs by testing a diverse range of quantitative methods, while not ignoring qualitative issues pertaining to the semantics-pragmatics interface. Starting from a critical assessment of what constitutes the meaning of modal verbs, different types of empirical studies (usage-based, data-driven and experimental), drawing considerably on the same data sets, shows how method triangulation can contribute to an enhanced understanding. Due attention is also given to individual variation as well as the degree to which modals can predict L2 proficiency level.

The Data Provenance Initiative: A Large Scale Audit of Dataset Licensing & Attribution in AI 37

The race to train language models on vast, diverse, and inconsistently documented datasets has raised pressing concerns about the legal and ethical risks for practitioners. To remedy these practices threatening data transparency and understanding, we convene a multidisciplinary effort between legal and machine learning experts to systematically audit and trace 1800+ text datasets. We develop tools and standards to trace the lineage of these datasets, from their source, creators, series of license conditions, properties, and subsequent use. Our landscape analysis highlights the sharp divides in composition and focus of commercially open vs closed datasets, with closed datasets monopolizing important categories: lower resource languages, more creative tasks, richer topic variety, newer and more synthetic training data. This points to a deepening divide in the types of data that are made available under different license conditions, and heightened implications for jurisdictional legal interpretations of copyright and fair use. We also observe frequent miscategorization of licenses on widely used dataset hosting sites, with license omission of 70%+ and error rates of 50%+. This points to a crisis in misattribution and informed use of the most popular datasets driving many recent breakthroughs. As a contribution to ongoing improvements in dataset transparency and responsible use, we release our entire audit, with an interactive UI, the Data Provenance Explorer, which allows practitioners to trace and filter on data provenance for the most popular open source finetuning data collections: www.dataprovenance.org.

Refined Convergence and Topology Learning for Decentralized SGD with Heterogeneous Data 20

One of the key challenges in decentralized and federated learning is to design algorithms that efficiently deal with highly heterogeneous data distributions across agents. In this paper, we revisit the analysis of the popular Decentralized Stochastic Gradient Descent algorithm (D-SGD) under data heterogeneity. We exhibit the key role played by a new quantity, called neighborhood heterogeneity, on the convergence rate of D-SGD. By coupling the communication topology and the heterogeneity, our analysis sheds light on the poorly understood interplay between these two concepts. We then argue that neighborhood heterogeneity provides a natural criterion to learn data-dependent topologies that reduce (and can even eliminate) the otherwise detrimental effect of data heterogeneity on the convergence time of D-SGD. For the important case of classification with label skew, we formulate the problem of learning such a good topology as a tractable optimization problem that we solve with a Frank-Wolfe algorithm. As illustrated over a set of simulated and real-world experiments, our approach provides a principled way to design a sparse topology that balances the convergence speed and the per-iteration communication costs of D-SGD under data heterogeneity.

One-Shot Federated Conformal Prediction 22

In this paper, we introduce a conformal prediction method to construct prediction sets in a oneshot federated learning setting. More specifically, we define a quantile-of-quantiles estimator and prove that for any distribution, it is possible to output prediction sets with desired coverage in only one round of communication. To mitigate privacy issues, we also describe a locally differentially private version of our estimator. Finally, over a wide range of experiments, we show that our method returns prediction sets with coverage and length very similar to those obtained in a centralized setting. Overall, these results demonstrate that our method is particularly well-suited to perform conformal predictions in a one-shot federated learning setting.

Differentially private speaker anonymization 18

Sharing real-world speech utterances is key to the training and deployment of voice-based services. However, it also raises privacy risks as speech contains a wealth of personal data. Speaker anonymization aims to remove speaker information from a speech utterance while leaving its linguistic and prosodic attributes intact. State-of-the-art techniques operate by disentangling the speaker information (represented via a speaker embedding) from these attributes and re-synthesizing speech based on the speaker embedding of another speaker. Prior research in the privacy community has shown that anonymization often provides brittle privacy protection, even less so any provable guarantee. In this work, we show that disentanglement is indeed not perfect: linguistic and prosodic attributes still contain speaker information. We remove speaker information from these attributes by introducing differentially private feature extractors based on an autoencoder and an automatic speech recognizer, respectively, trained using noise layers. We plug these extractors in the state-of-the-art anonymization pipeline and generate, for the first time, differentially private utterances with a provable upper bound on the speaker information they contain. We evaluate empirically the privacy and utility resulting from our differentially private speaker anonymization approach on the LibriSpeech data set. Experimental results show that the generated utterances retain very high utility for automatic speech recognition training and inference, while being much better protected against strong adversaries who leverage the full knowledge of the anonymization process to try to infer the speaker identity.

From Noisy Fixed-Point Iterations to Private ADMM for Centralized and Federated Learning 21

We study differentially private (DP) machine learning algorithms as instances of noisy fixed-point iterations, in order to derive privacy and utility results from this well-studied framework. We show that this new perspective recovers popular private gradient-based methods like DP-SGD and provides a principled way to design and analyze new private optimization algorithms in a flexible manner. Focusing on the widely-used Alternating Directions Method of Multipliers (ADMM) method, we use our general framework to derive novel private ADMM algorithms for centralized, federated and fully decentralized learning. For these three algorithms, we establish strong privacy guarantees leveraging privacy amplification by iteration and by subsampling. Finally, we provide utility guarantees using a unified analysis that exploits a recent linear convergence result for noisy fixed-point iterations.

DP-SGD Without Clipping: The Lipschitz Neural Network Way 36

State-of-the-art approaches for training Differentially Private (DP) Deep Neural Networks (DNN) faces difficulties to estimate tight bounds on the sensitivity of the network's layers, and instead rely on a process of per-sample gradient clipping. This clipping process not only biases the direction of gradients but also proves costly both in memory consumption and in computation. To provide sensitivity bounds and bypass the drawbacks of the clipping process, our theoretical analysis of Lipschitz constrained networks reveals an unexplored link between the Lipschitz constant with respect to their input and the one with respect to their parameters. By bounding the Lipschitz constant of each layer with respect to its parameters we guarantee DP training of these networks. This analysis not only allows the computation of the aforementioned sensitivities at scale but also provides leads on to how maximize the gradient-to-noise ratio for fixed privacy guarantees. To facilitate the application of Lipschitz networks and foster robust and certifiable learning under privacy guarantees, we provide a Python package that implements building blocks allowing the construction and private training of such networks.

GAP: Differentially Private Graph Neural Networks with Aggregation Perturbation 30

In this paper, we study the problem of learning Graph Neural Networks (GNNs) with Differential Privacy (DP). We propose a novel differentially private GNN based on Aggregation Perturbation (GAP), which adds stochastic noise to the GNN's aggregation function to statistically obfuscate the presence of a single edge (edge-level privacy) or a single node and all its adjacent edges (node-level privacy). Tailored to the specifics of private learning, GAP's new architecture is composed of three separate modules: (i) the encoder module, where we learn private node embeddings without relying on the edge information; (ii) the aggregation module, where we compute noisy aggregated node embeddings based on the graph structure; and (iii) the classification module, where we train a neural network on the private aggregations for node classification without further querying the graph edges. GAP's major advantage over previous approaches is that it can benefit from multi-hop neighborhood aggregations, and guarantees both edge-level and node-level DP not only for training, but also at inference with no additional costs beyond the training's privacy budget. We analyze GAP's formal privacy guarantees using Rényi DP and conduct empirical experiments over three real-world graph datasets. We demonstrate that GAP offers significantly better accuracy-privacy trade-offs than state-of-the-art DP-GNN approaches and naive MLP-based baselines. Our code is publicly available here.

High-Dimensional Private Empirical Risk Minimization by Greedy Coordinate Descent 25

In this paper, we study differentially private empirical risk minimization (DP-ERM). It has been shown that the worst-case utility of DP-ERM reduces polynomially as the dimension increases. This is a major obstacle to privately learning large machine learning models. In high dimension, it is common for some model's parameters to carry more information than others. To exploit this, we propose a differentially private greedy coordinate descent (DP-GCD) algorithm. At each iteration, DP-GCD privately performs a coordinate-wise gradient step along the gradients' (approximately) greatest entry. We show theoretically that DP-GCD can achieve a logarithmic dependence on the dimension for a wide range of problems by naturally exploiting their structural properties (such as quasi-sparse solutions). We illustrate this behavior numerically, both on synthetic and real datasets.

Differential Privacy has Bounded Impact on Fairness in Classification 26

We theoretically study the impact of differential privacy on fairness in classification. We prove that, given a class of models, popular group fairness measures are pointwise Lipschitz-continuous with respect to the parameters of the model. This result is a consequence of a more general statement on accuracy conditioned on an arbitrary event (such as membership to a sensitive group), which may be of independent interest. We use this Lipschitz property to prove a non-asymptotic bound showing that, as the number of samples increases, the fairness level of private models gets closer to the one of their non-private counterparts. This bound also highlights the importance of the confidence margin of a model on the disparate impact of differential privacy.

Rényi Pufferfish Privacy: General Additive Noise Mechanisms and Privacy Amplification by Iteration via Shift Reduction Lemmas 38

Pufferfish privacy is a flexible generalization of differential privacy that allows to model arbitrary secrets and adversary's prior knowledge about the data. Unfortunately, designing general and tractable Pufferfish mechanisms that do not compromise utility is challenging. Furthermore, this framework does not provide the composition guarantees needed for a direct use in iterative machine learning algorithms. To mitigate these issues, we introduce a Rényi divergence-based variant of Pufferfish and show that it allows us to extend the applicability of the Pufferfish framework. We first generalize the Wasserstein mechanism to cover a wide range of noise distributions and introduce several ways to improve its utility. We also derive stronger guarantees against out-ofdistribution adversaries. Finally, as an alternative to composition, we prove privacy amplification results for contractive noisy iterations and showcase the first use of Pufferfish in private convex optimization. A common ingredient underlying our results is the use and extension of shift reduction lemmas.

Exploiting Problem Structure in Privacy-Preserving Optimization and Machine Learning 33

In the past decades, concerns about the societal impact of machine learning have been growing. Indeed, if machine learning has proven its usefulness in science, day-to-day applications, and many other domains, its success is principally due to the availability of large datasets. This raises two concerns, the first about the confidentiality of the training data, and the second, about possible discrimination in a model's predictions. Trustworthy machine learning aims at providing technical answers to these concerns.Unfortunately, guaranteeing the privacy of the training data and the fairness of the predictions often decreases the utility of the learned model. This problem has drawn significant interest in the past years, but most of existing methods (usually based on stochastic gradient descent) tend to fail in some common scenarios, like training of high-dimensional models. In this thesis, we study how structural properties of machine learning problems can be exploited to improve the trade-off between privacy and utility, and how this can impact the fairness of the predictions.The first two contributions of this thesis are two new differentially private optimization algorithms, that are both based on coordinate descent. They aim at exploiting different structural properties of the problem at hand. The first algorithm is based on stochastic coordinate descent, and can exploit imbalance in the scale of the gradient's coordinates by using large step sizes. This allows our algorithm to obtain useful models in difficult problems, where stochastic gradient descent quickly stalls. The second algorithm is based on greedy coordinate descent. Its greedy updates allow to focus on the most important coordinates of the problem, which can sometimes drastically improve utility (e.g. when the solution of the problem is sparse).The third contribution of this thesis studies the interplay of differential privacy and fairness in machine learning. These two notions have rarely been studied simultaneously, and there are growing concerns that differential privacy may exacerbate unfairness. We show that group fairness measures have interesting regularity properties, provided that the predictions of the model are Lipschitz-continuous in its parameters. This result allows to derive a bound on the difference in fairness levels between a private model and its non-private counterpart.

Privacy-preserving Learning by Averaging in Collaborative Networks 34

In recent years, due to the growing importance of network applications and the growing concerns for privacy, there is an increasing interest in decentralized forms of machine learning. In this dissertation, we study the setting that involves a communication network of agents, where each agent locally privatizes (adds noise to) its data, and where the agents aim to collaboratively learn statistical models over their data. Such local privatization is in line with a standard of data privacy known as local differential privacy, and local differential privacy is useful when alternatives, such as secure multi-party computation or central differential privacy performed by a trusted curator, are infeasible. However, local differential privacy results, typically, in worse utility (less accurate statistical models) compared to central differential privacy because, for the same privacy budget, local differential privacy adds more privatization noise than central differential privacy. The principal question of this dissertation is the following: given that local differential privacy must be used, how could the agents maximize the utility they achieve? We study two cases to address the stated principal question.In the first case, we consider the problem of distributed averaging, where each agent intends to collaboratively compute the unbiased average over the individual values of all agents without revealing neither their sensitive attributes nor their degree (number of neighbors). Usually, existing works solve this problem by assuming that either (i) each agent reveals its degree to its neighbors or (ii) every two neighboring agents can perform handshakes (requests that rely on replies) in every exchange of information. Since such assumptions are not always desirable, we propose an approach that is handshake-free and where the degrees are privatized. In particular, we use a gossip algorithm that computes averages that are biased when the graph of agents is non-regular (when the vertices have unequal degrees) and then perform a procedure combining multiple biased averages for bias correction. We apply the proposed approach for fitting linear regression models. We prove the asymptotic guarantee that the mean squared error between the average of privatized attributes computed by our approach and the average of sensitive attributes is $𝒪\left(\frac{1}{n}\right)$, where $n$ is the number of agents.In the second case, we consider a group of agents, where features (for fitting regression models) are computed by transforming sensitive attributes, and where the transformations have high-magnitude gradients or singularities. In such setting, there is a risk to magnify the privatization noise if the perturbed data is in an interval where the feature function has high-magnitude gradients. We provide a tailored noise mechanism for privatizing features by solving a convex program in such a way that (i) only pertinent intervals of transformations are selected, (ii) the variance of privatization noise is minimized, and (iii) the biasedness of privatization noise is minimized.

Introducing the TRUMPET project: TRUstworthy Multi-site Privacy Enhancing Technologies 19

This paper is an overview of the EU-funded project TRUMPET (see here), and gives an outline of its scope and main technical aspects and objectives. In recent years, Federated Learning has emerged as a revolutionary privacy-enhancing technology. However, further research has cast a shadow of doubt on its strength for privacy protection. The goal of TRUMPET is to research and develop novel privacy enhancement methods for Federated Learning, and to deliver a highly scalable Federated AI service platform for researchers, that will enable AI-powered studies of siloed, multi-site, crossdomain, cross-border European datasets with privacy guarantees that follow the requirements of GDPR. The generic TRUMPET platform will be piloted, demonstrated and validated in the specific use case of European cancer hospitals, allowing researchers and policymakers to extract AI-driven insights from previously inaccessible cross-border, cross-organization cancer data, while ensuring the patients' privacy.

Fair Without Leveling Down: A New Intersectional Fairness Definition 24

In this work, we consider the problem of intersectional group fairness in the classification setting, where the objective is to learn discrimination-free models in the presence of several intersecting sensitive groups. First, we illustrate various shortcomings of existing fairness measures commonly used to capture intersectional fairness. Then, we propose a new definition called the $\alpha$-Intersectional Fairness, which combines the absolute and the relative performance across sensitive groups and can be seen as a generalization of the notion of differential fairness. We highlight several desirable properties of the proposed definition and analyze its relation to other fairness measures. Finally, we benchmark multiple popular in-processing fair machine learning approaches using our new fairness definition and show that they do not achieve any improvement over a simple baseline. Our results reveal that the increase in fairness measured by previous definitions hides a ”leveling down” effect, i.e., degrading the best performance over groups rather than improving the worst one.

FairGrad: Fairness Aware Gradient Descent 14

We tackle the problem of group fairness in classification, where the objective is to learn models that do not unjustly discriminate against subgroups of the population. Most existing approaches are limited to simple binary tasks or involve difficult to implement training mechanisms. This reduces their practical applicability. In this paper, we propose FairGrad, a method to enforce fairness based on a reweighting scheme that iteratively learns group specific weights based on whether they are advantaged or not. FairGrad is easy to implement and can accommodate various standard fairness definitions. Furthermore, we show that it is comparable to standard baselines over various datasets including ones used in natural language processing and computer vision.

Private Sampling with Identifiable Cheaters 16

In this paper we study verifiable sampling from probability distributions in the context of multi-party computation. This has various applications in randomized algorithms performed collaboratively by parties not trusting each other. One example is differentially private machine learning where noise should be drawn, typically from a Laplace or Gaussian distribution, and it is desirable that no party can bias this process. In particular, we propose algorithms to draw random numbers from uniform, Laplace, Gaussian and arbitrary probability distributions, and to verify honest execution of the protocols through zero-knowledge proofs. We propose protocols that result in one party knowing the drawn number and protocols that deliver the drawn random number as a shared secret.

A Revenue Function for Comparison-Based Hierarchical Clustering 15

Comparison-based learning addresses the problem of learning when, instead of explicit features or pairwise similarities, one only has access to comparisons of the form: Object $A$ is more similar to $B$ than to $C$. Recently, it has been shown that, in Hierarchical Clustering, single and complete linkage can be directly implemented using only such comparisons while several algorithms have been proposed to emulate the behaviour of average linkage. Hence, finding hierarchies (or dendrograms) using only comparisons is a well understood problem. However, evaluating their meaningfulness when no ground-truth nor explicit similarities are available remains an open question. In this paper, we bridge this gap by proposing a new revenue function that allows one to measure the goodness of dendrograms using only comparisons. We show that this function is closely related to Dasgupta's cost for hierarchical clustering that uses pairwise similarities. On the theoretical side, we use the proposed revenue function to resolve the open problem of whether one can approximately recover a latent hierarchy using few triplet comparisons. On the practical side, we present principled algorithms for comparison-based hierarchical clustering based on the maximisation of the revenue and we empirically compare them with existing methods.

Automated detection of toxicophores and prediction of mutagenicity using PMCSFG algorithm 17

Maximum common substructures (MCS) have received a lot of attention in the chemoinformatics community. They are typically used as a similarity measure between molecules, showing high predictive performance when used in classification tasks, while being easily explainable substructures. In the present work, we applied the Pairwise Maximum Common Subgraph Feature Generation (PMCSFG) algorithm to automatically detect toxicophores (structural alerts) and to compute fingerprints based on MCS. We present a comparison between our MCS-based fingerprints and 12 well-known chemical fingerprints when used as features in machine learning models. We provide an experimental evaluation and discuss the usefulness of the different methods on mutagenicity data. The features generated by the MCS method have a state-of-the-art performance when predicting mutagenicity, while they are more interpretable than the traditional chemical fingerprints.

A Smartphone-based Architecture for Prolonged Monitoring of Gait 35

Gait analysis is important for evaluating neurological disorders such as stroke and Parkinson's disease. Traditionally, healthcare professionals had to rely on subjective assessments (i.e., human-based) of gait which were time consuming and not very reproducible. However, with the advent of IoT and indeed more objective (e.g., measurement-based) assessment methods, gait analysis can now be performed more accurately and effectively. It is worth noting, however, that there are still limitations to these objective methods, especially the lack of privacy-preserving continuous data collection. To overcome this limitation, we present in this paper a privacy-by-design monitoring application for post-stroke patients to evaluate their gait before, during, and after a rehabilitation program. Gait measurements are collected by a mobile application that continuously captures spatiotemporal parameters in the background using the built-in smartphone accelerometer. Statistical techniques are then applied to extract general indicators about the performed activity, as well as some more specific gait metrics in real-time such as regularity, symmetry and walking speed. These metrics are calculated based on the detected steps while patients are performing an activity. Additionally, a deep learning approach based on an auto-encoder is implemented to detect abnormal activities in the gait of patients. These analyses provides both valuable insights and statistical information about the activities performed by the patient, and a useful tool for practitioners to monitor the progression of neurological disorders and detect anomalies. We conducted experiments using this application in real conditions to monitor post-stroke patients in collaboration with a hospital, demonstrating its ability to compute valuable metrics and detect abnormal events patient's gait.

Linear Programs with Conjunctive Database Queries 13

In this paper, we study the problem of optimizing a linear program whose variables are the answers to a conjunctive query. For this we propose the language LP(CQ) for specifying linear programs whose constraints and objective functions depend on the answer sets of conjunctive queries. We contribute an efficient algorithm for solving programs in a fragment of LP(CQ). The natural approach constructs a linear program having as many variables as there are elements in the answer set of the queries. Our approach constructs a linear program having the same optimal value but fewer variables. This is done by exploiting the structure of the conjunctive queries using generalized hypertree decompositions of small width to factorize elements of the answer set together. We illustrate the various applications of LP(CQ) programs on three examples: optimizing deliveries of resources, minimizing noise for differential privacy, and computing the s-measure of patterns in graphs as needed for data mining.

Transfer learning for text anonymization

Participants: Damien Siléo, Marc Tommasi, Gabriel Loiseau.

VADE is a major company that processes emails at large scale to detect attacks like phishing.

In this project we design utility and privacy evaluation methods based on the combination of many tasks and objectives, relevant in the text (email) context. We study and compare approaches based on text generation or based on the replacement or obfuscation of selected entities, to tune the privacy utility trade-off.

Synthetic data generation with privacy constraints

Participants: Aurélien Bellet, Marc Tommasi, Clément Pierquin.

Craft.ai is a company whose activity was originally focused on explainable models for time series. It offers now MLops solutions based on AI with trustworthy guarantees. In this bilateral project with Craft.ai, Magnet brings expertise in privacy preserving machine learning for the generation of synthetic data.

The project is organized in four major axes. The definition of quality metrics for synthetic data; the design of algorithms for synthetic data generation with differential privacy guarantees; the definition of theoretical and empirical bounds on privacy associated with the release of synthetic data sets or generative models; some applications on time series or correlated data.

10.1.1 Participation in other International Programs

10.2.1 Visits to international teams

10.3.1 Horizon Europe

10.4.1 ANR DEEP-Privacy (2019–2023)

Participants: Marc Tommasi [contact person], Aurélien Bellet, Pascal Denis, Jan Ramon, Brij Mohan Lal Srivastava, Rishabh Gupta.

DEEP-PRIVACY proposes a new paradigm based on a distributed, personalized, and privacy-preserving approach for speech processing, with a focus on machine learning algorithms for speech recognition. To this end, we propose to rely on a hybrid approach: the device of each user does not share its raw speech data and runs some private computations locally, while some cross-user computations are done by communicating through a server (or a peer-to-peer network). To satisfy privacy requirements at the acoustic level, the information communicated to the server should not expose sensitive speaker information.

10.4.2 HyAIAI. INRIA Defi (2019-2023)

Participants: Jan Ramon [contact person], Marc Tommasi.

HyAIAI is an Inria Defi about the design of novel, interpretable approaches for Artificial Intelligence.

Recent progress in Machine Learning (ML) and especially Deep Learning has made ML pervasive in a wide range of applications. However, current approaches rely on complex numerical models: their decisions, as accurate as they may be, cannot be easily explained to the layman that may depend on these decisions (ex: get a loan or not). In the HyAIAI IPL, we tackle the problem of making “Interpretable ML” through the study and design of hybrid approaches that combine state of the art numeric models with explainable symbolic models. More precisely, our goal is to be able to integrate high level (domain) constraints in ML models, to give model designers information on ill-performing parts of the model, and to give the layman/practitioner understandable explanations on the results of the ML model.

10.4.3 ANR PMR (2020-2024)

Participants: Jan Ramon [contact person], Aurélien Bellet, Marc Tommasi, Cesar Sabater.

Given the growing awareness of privacy risks of data processing, there is an increasing interest in privacy-preserving learning. However, shortcomings in the state of the art limit the applicability of the privacy-preserving learning paradigm. First, most approaches assume too optimistically a honest-but-curious setting. Second, most approaches consider one learning task in isolation, not accounting for the context where querying is a recurring activity. We will investigate new algorithms and models that address these shortcomings. Among others, (i) our algorithms will combine privacy-preserving properties of differential privacy with security offered by cryptography and (ii) based on models of information flows in integrated data handling processes, we will build more refined models analyzing the implications of repeated querying. We will demonstrate the utility of our new theory and algorithms by proposing strategies to realistically apply them in significant real-world problems illustrated through use cases in the medical domain

10.4.4 FedMalin. INRIA Defi (2021-2024)

Participants: Aurélien Bellet [contact person], Jan Ramon, Marc Tommasi, Michaël Perrot, Batiste Le Bars, Edwige Cyffers, Paul Mangold, Tudor Cebere.

In many use-cases of Machine Learning (ML), data is naturally decentralized: medical data is collected and stored by different hospitals, crowdsensed data is generated by personal devices, etc. Federated Learning (FL) has recently emerged as a novel paradigm where a set of entities with local datasets collaboratively train ML models while keeping their data decentralized.

FedMalin is a research project that spans 10 Inria research teams and aims to push FL research and concrete use-cases through a multidisciplinary consortium involving expertise in ML, distributed systems, privacy and security, networks, and medicine. We propose to address a number of challenges that arise when FL is deployed over the Internet, including privacy and fairness, energy consumption, personalization, and location/time dependencies.

FedMalin will also contribute to the development of open-source tools for FL experimentation and real-world deployments, and use them for concrete applications in medicine and crowdsensing.

10.4.5 FLAMED: Federated Learning and Analytics on Medical Data. INRIA Action Exploratoire (2020-2024)

Participants: Aurélien Bellet [contact person], Marc Tommasi, Paul Mangold, Nathan Bigaud, Paul André.

Flamed is about decentralized approaches for AI in medicine. The main objective is to operate data analysis and machine learning tasks in a network of hospital units without any data exchange. This approach helps to solve data privacy and sovereignty issues while taking advantage of the statistical power of federation and collaboration. This research is done in collaboration with the Lille Hospital.

10.4.6 ANR-JCJC PRIDE (2020–2025)

Participants: Aurélien Bellet [contact person], Marc Tommasi, Jan Ramon, Edwige Cyffers, Batiste Le Bars, Paul Mangold, Tudor Cebere.

Machine learning (ML) is ubiquitous in AI-based services and data-oriented scientific fields but raises serious privacy concerns when training on personal data. The starting point of PRIDE is that personal data should belong to the individual who produces it. This requires to revisit ML algorithms to learn from many decentralized personal datasets while preventing the reconstruction of raw data. Differential Privacy (DP) provides a strong notion of protection, but current decentralized ML algorithms are not able to learn useful models under DP. The goal of PRIDE is to develop theoretical and algorithmic tools that enable differentially-private ML methods operating on decentralized datasets, through two complementary objectives: (1) prove that gossip protocols naturally reinforce DP guarantees; (2) propose algorithms at the intersection of decentralized ML and secure multi-party computation.

10.4.7 COMANCHE: Computational Models of Lexical Meaning and Change. INRIA Action Exploratoire (2022-2026)

Participants: Pascal Denis [contact person], Mikaela Keller, Bastien Liétard.

Comanche proposes to transfer and adapt recent Natural Language representation learning algorithms from deep learning to model the evolution of the meaning of words, and to confront these computational models to theories on language acquisition and the diachrony of languages. At the crossroads between machine learning, psycholinguistics and historical linguistics, this project will make it possible to validate or revise some of these theories, but also to bring out computational models that are more sober in terms of data and computations because they exploit new inductive biases inspired by these disciplines.

In collaboration with UMR SCALAB (CNRS, Université de Lille), l’Unité de Recherche STIH (Sorbonne Université), et l’UMR ATILF (CNRS, Université de Lorraine).

10.4.8 IPoP, Projet interdisciplinaire sur la protection des données personnelles, PEPR Cybersécurité (2022-2028).

Participants: Aurélien Bellet [contact person], Jan Ramon, Marc Tommasi, Michaël Perrot, Cesar Sabater, Edwige Cyffers, Paul Mangold, Tudor Cebere.

Digital technologies provide services which can greatly increase quality of life (e.g. connected e-health devices, location based services, or personal assistants). However, these services can also raise major privacy risks, as they involve personal data, or even sensitive data. Indeed, this notion of personal data is the cornerstone of French and European regulations, since processing such data triggers a series of obligations that the data controller must abide by. This raises many multidisciplinary issues, as the challenges are not only technological, but also societal, judiciary, economic, political and ethical.

The objectives of this project are thus to study the threats on privacy that have been introduced by these new services, and to conceive theoretical and technical privacy-preserving solutions that are compatible with French and European regulations, that preserve the quality of experience of the users. These solutions will be deployed and assessed, both on the technological and legal sides, and on their societal acceptability. In order to achieve these objectives, we adopt an interdisciplinary approach, bringing together many diverse fields: computer science, technology, engineering, social sciences, economy and law.

The project's scientific program focuses on new forms of personal information collection, on Artificial Intelligence (AI) and its governance, data anonymization techniques, personal data management and distributed calculation protocol privacy preserving infrastructures, differential privacy, personal data legal protection and compliance, and all the associated societal and ethical considerations. This unifying interdisciplinary research program brings together internationally recognized research teams (from universities, engineering schools and institutions) working on privacy, and the French Data Protection Authority (CNIL).

This holistic vision of the issues linked to personal data protection will on the one hand let us propose solutions to the scientific and technological challenges and on the other help us confront these solutions in many different ways, in the context of interdisciplinary collaborations, thus leading to recommendations and proposals in the field of regulations or legal frameworks. This comprehensive consideration of all the issues aims at encouraging the adoption and acceptability of the solutions proposed by all stakeholders, legislators, data controllers, data processors, solution designers, developers all the way to end-users.

10.4.9 CAPS'UL (2023-2028)

Participant: Marc Tommasi [contact person].

The project is built around 3 axes.

1. Promote a common digital health culture among all current and future healthcare professionals: cybersecurity issues, legal and ethical regulation of healthcare data, communication and digital health tools, telehealth framework.
2. Design a high-performance tool for practical situations, enabling concrete and effective collaboration between the various training, socio-economic and medico-social players in the implementation of training courses. This shared resource center will provide a credible immersive environment (real software and simulated healthcare data) and teaching scenarios for the entire teaching community. Co-constructed with industry software publishers, it will be accessible from simulation centers and remotely, to meet the different needs of the region.
3. Train professionals in the new digital health support professions, by emphasizing the delivery of “health and specific digital issues” courses that are shared between the various existing courses. These innovative, coherent schemes will serve as demonstrators of excellence on a regional scale.

Magnet will provide tools for synthetic data generation with privacy guarantees dedicated to the immersive environment.

10.4.10 ANR-JCJC FaCTor: Fairness Constraints and Guarantees for Trustworthy Machine Learning (2023-2027)

Participants: Michaël Perrot [contact person], Marc Tommasi.

The goal of the FaCTor project is to provide ML practitioners with theoretically well founded means to develop algorithms that come with fairness guarantees. It points toward the development of trustworthy and socially acceptable ML solutions. The end goal is to make the models more accountable and in line with the requirements of the law, ensuring that the benefits of ML are not limited to a subset of the population.

10.4.11 REDEEM: Resilient, Decentralized and Privacy-Preserving Machine Learning, PEPR IA (2022-2028).

Participants: Jan Ramon [contact person], Marc Tommasi, Michaël Perrot, Arnaud Descours, Vitalii Emilianov.

The vision of distributed AI is attractive because it contributes to user empowerment by limiting the dissemination of personal and confidential information to a single node in the network and it makes systems independent of a superior force that would decide what is good for everyone. But on the other hand it opens up major issues of security and robustness: how can we guarantee the compliance of a model learned in another context? How can we protect our AI network from the introduction of biased knowledge, malicious or not, or even “backdoor” functions? If the pooling consists of a simultaneous optimisation, how can we ensure the validity of contributions that are not always explicable?

The action led on the theme of distributed AI is therefore at the confluence of the topics Embedded and Frugality (distributed systems are frequently low-resource embedded systems such as telephones, vehicles or autonomous robots) and Trust, as the issues of security, reliability and robustness are shed in a new light in collaborative AI.

The REDEEM project brings together a consortium of complementary teams and researchers, with primary expertise in machine learning, distributed optimization, consensus algorithms and game theory. It also associates a unique spectrum of research orientation, from highly theoretical work on convergence of distributed learning algorithms to extensive experiences towards practical and efficient implementations as well as innovative dissemination activities.

10.5.1 STARS: Fairness in decentralized and privacy-preserving machine learning (2021-2023).

Participant: Michaël Perrot [contact person].

The aim of this project is to propose approaches capable of learning a non-discriminatory model in a context where data is distributed among several entities that wish to preserve the confidentiality of the data in their possession.

11.1.1 Scientific events: organisation

11.1.2 Scientific events: selection

• Marc Tommasi served as Area Chair for ECML-PKDD and ICLR and PC member for UAI (top reviewer).
• Aurélien Bellet served as Area Chair for ICML and NeurIPS.
• Michaël Perrot served as Reviewer for NeurIPS (Top Reviewer) and ICML
• Jan Ramon was PC member of AFT@NeurIps'23, AISTATS'24, BNAIC'23, DGM4G@NeurIps'23, DS'23, ECML/PKDD'23, ICBINB@NeurIps'23, ICLR'24, ICML'23, IJCAI'23, MLG@ECML/PKDD'23, MLG@KDD'23, NeurIps'23, PPAI@AAAI'23, SDM'24, XKDD@ECML/PKDD'23.
• Mikaela Keller served as Reviewer for *SEM'23 and CAp'23.
• Damien Sileo served as Reviewer for ACL Rolling Reviews.
• Pascal Denis was Action Editor for ACL Rolling Review, and served as PC member for ACL'23, EACL'23, and CODI@ACL'23.

11.1.3 Journal

• Aurélien Bellet is Action Editor for Transactions of Machine Learning Research (TMLR).
• Jan Ramon is member of the editorial boards of Machine Learning Journal (MLJ), Data Mining and Knowledge Discovery (DMKD), Journal of Machine Learning Research (JMLR), ECML-PKDD Journal track. Jan Ramon is action editor of Data Mining and Knowledge Discovery (DMKD).
• Pascal Denis is standing reviewer for Transactions of the Association for Computational Linguistics (TACL).

11.1.4 Invited talks

• Michaël Perrot gave a talk titled “Differential Privacy has Bounded Impact on Fairness in Classification” at the MICS Seminars.
• Damien Sileo gave a talk named “Generative AI and Large Language Models” at the Cyle Supérieur du Numérique.
• Aurélien Bellet gave a talk at the Privacy and Fairness in AI for Health workshop, the Paris Privacy-Preserving ML workshop and the Privacy Alpine Seminar (Privaski).
• Marc Tommasi gave a talk at the days on Distributed Learning organized by the GDR RSD in association with SIF.

11.1.5 Scientific expertise

• Marc Tommasi was a member (scientific expert) of the recruitment committee of full professors at Lens and Marseille.
• Aurélien Bellet was a member of the recruitement committee of an assistant professor at Université Jean Monnet de Saint-Etienne.
• Mikaela Keller was a member of the recruitement committees of an assistant professor at Université Jean Monnet de Saint-Etienne and of an assistant professor at Université de Lille.
• Aurélien Bellet serves as ethics advisor for the ESFRI project SLICES-PP.
• Jan Ramon was reviewer for ANR, COST, Horizon Europe and ChistERA.

11.1.6 Research administration

• Jan Ramon was a member of CER (Commission Emploi Recherche) in the INRIA Center of Lille University.
• Aurélien Bellet is member of the Operational Committee for the assessment of Legal and Ethical risks (COERLE).
• Marc Tommasi is co-head of the DatInG group (4 teams, about 100 persons), member of the Conseil Scientifique du laboratoire CRIStAL and member of the Commission mixte CRIStAL/Faculty of Science, Lille University.
• Pascal Denis is co-head of the CNRS GDR "Langues et langage à la croisée des disciplines" (LLcD) and a member of the CNRS GDR NLP Group. Pascal Denis is also a member of the network "référents données" at Inria and Université de Lille (Lille Open Research Data). He is administrator of Inria membership to Linguistic Data Consortium (LDC).

11.2.1 Teaching

• Licence MIASHS: Marc Tommasi, Data Science, 24h, L2, Université de Lille.
• Licence Informatique: Marc Tommasi, Introduction to AI, 24h, L2, Université de Lille.
• Licence MIASHS: Mikaela Keller, Data Science, 12h, L2, Université de Lille.
• Licence MIASHS: Mikaela Keller, Data Science 2, 24h, L3, Université de Lille.
• Licence MIASHS: Mikaela Keller, Traitement de données, 24h, L2, Université de Lille.
• Master MIASHS: Mikaela Keller, Algorithmes fondamentaux de la fouille de données, 27h, M1, Université de Lille.
• Master Data Science: Mikaela Keller, Machine Learning 1, 24h, M1, Université de Lille.
• Master Computer Science: Mikaela Keller, Apprentissage profond, 24h, M1, Université de Lille.
• Master Computer Science: Mikaela Keller, Machine learning pour le traitement automatique du language naturel, 24h, M2, Université de Lille.
• Master MIASHS: Michaël Perrot, Algorithmes fondamentaux de la fouille de données, 27h, M1, Université de Lille.
• Master Computer Science: Marc Tommasi, Data Science, 48h, M1, Université de Lille.
• Master Computer Science: Marc Tommasi, Semi-supervised learning and Graphs, 24h, M2, Université de Lille.
• Master Data Science: Marc Tommasi Seminars 24h.
• Master Data Science: Aurélien Bellet, Privacy Preserving Machine Learning, 24h, M2, Université de Lille and Ecole Centrale de Lille.
• Parcours Science des Données et Intelligence Artificielle: Aurélien Bellet, Advanced Machine Learning, 6h, Ecole Centrale de Lille.
• Training in Privacy-Preserving and Federated Machine Learning, 3 days, researchers from L'Oréal R&D.
• Master Sciences Cognitives: Damien Sileo, Machine Learning for Cognitive Sciences, 8h, M2, Université de Lille.
• Master Data Science: Michaël Perrot, Fairness in Trustworthy Machine Learning, 24h, M2, Université de Lille et Ecole Centrale de Lille.
• Marc Tommasi is directeur des études for the Machine Learning master of Computer Science.

11.2.2 Supervision

• Postdoc: Vitalii Emilianov. On the interactions between fairness and privacy in machine learning. Since Nov. 2022. Michaël Perrot.
• Postdoc: Batiste Le Bars. On collaboration graph design for decentralized learning. October 2021-Jun. 2023. Aurélien Bellet and Marc Tommasi
• Postdoc: Arnaud Descours. On federated optimization with lower communication cost. Since Nov. 2023. Jan Ramon
• Postdoc: Baptiste Cottier. On sparse secure aggregation. May-Sept 2023. Jan Ramon
• Postdoc: Sabri Chellouf. On privacy-preserving learning for mobile devices. May 2023 - Dec 2023. Jan Ramon
• Postdoc: Imane Talbi. On the relation between statistical privacy and security assumptions. May 2023-April 2024. Jan Ramon
• PhD defended in Feb. 2023: Nicolas Crosetti, Privacy Risks of Aggregates in Data Centric-Workflows. Florent Capelli and Sophie Tison and Joachim Niehren and Jan Ramon.
• Phd defended in Jun. 2023: Arijus Pleska, Tractable Probabilistic Models for Large Scale Networks 34. Jan Ramon
• PhD in progress: Moitree Basu, Integrated privacy-preserving AI, since 2019. Jan Ramon.
• PhD defended in Oct. 2023: Paul Mangold. Decentralized Optimization and privacy 33. Aurélien Bellet and Marc Tommasi and Joseph Salmon, since October 2020.
• Phd in progress: Gaurav Maheshwari. Trustworthy Representations for Natural Language Processing, since Nov 2020. Aurélien Bellet, Mikaela Keller, and Pascal Denis
• Phd in progress: Edwige Cyffers. Decentralized learning and privacy amplification, since Oct. 2021. Aurélien Bellet
• Phd in progress: Marc Damie. Secure protocols for verifiable decentralized machine learning, since May 2022. Jan Ramon with Andreas Peter (U. Twente, NL & U. Oldenburg, DE) Florian Hahn (University of Twente, NL).
• Phd in progress: Tudor Cebere. Privacy-Preserving Machine Learning, since Nov. 2022. Aurélien Bellet
• Phd in progress: Bastien Liétard. Computational Models of Lexical Semantic Change, since Nov. 2022. Anne Carlier (Université Paris Sorbonne), Pascal Denis and Mikaela Keller
• Phd in progress: Dinh-Viet-Toan Le. Natural Language Processing approaches in the musical domain : suitability, performance and limits, since Oct. 2022. Mikaela Keller and Louis Bigo
• Phd in progress: Aleksei Korneev. Trustworthy multi-site privacy-enhancingtechnologies, since Dec. 2022. Jan Ramon
• PhD in progress: Antoine Barczewski. Transparent privacy-preserving machine learning, since May 2022. Jan Ramon.
• PhD in progress: Aurélien Saïd Housseini. Computational Models of Semantic Memory, since Sept. 2023, Angèle Brunellière (UMR SCALab, Université de Lille) Pascal Denis and Rémi Gilleron.
• PhD in progress: Clément Pierquin Synthetic data generation with privacy constraints, since Sept. 2023, Aurélien Bellet and Marc Tommasi.
• PhD in progress: Gabriel Loiseau Transfert and multitask learning approaches for text anonymizaton, since Sept. 2023, Damien Sileo and Marc Tommasi.
• PhD in progress: Brahim Erraji Privacy and Fairness, since Sept. 2023, Aurélien Bellet, Catuscia Palamidessi and Michaël Perrot.
• Engineer Sophie Villerot, ADT project Tailed: Trustworthy AI Library for Environments which are Decentralized, since Nov. 2020. Jan Ramon
• Engineer Kamalkumar Macwan, ADT project Tailed: Trustworthy AI Library for Environments which are Decentralized, Jan-Oct. 2023. Jan Ramon
• Engineer Joseph Renner, Improving Word Representations with Semantic Knowledge, Nov. 2020-Oct 2023. Pascal Denis and Rémi Gilleron
• Engineer Rishabh Gupta, Disentanglement approaches for speech data. Apr. 2023-Jul. 2023. Aurélien Bellet and Marc Tommasi.
• Engineer Paul Andrey, Decentralized and Federated Learning with DecLearn. Since Jul. 2022. Aurélien Bellet and Marc Tommasi.
• Engineer Nathan Bigaud, Decentralized and Federated Learning with DecLearn. Oct. 2022-Oct.2023.Aurélien Bellet and Marc Tommasi.
• Engineer Quentin Sinh. Integrating MAGNET results in the TRUMPET privacy-preserving federated learning platform, Since Nov 2023. Jan Ramon.
• Engineer Kevin Ngakossa. Integrating MAGNET results on constraint-based privacy in the TRUMPET privacy-preserving federated learning platform, Since Oct. 2023. Jan Ramon.
• Engineer Léonard Deroose. Development of TRUMPET privacy-preserving platform components, Since Sep. 2023. Jan Ramon.
• Engineer Li Mou. Statistical privacy computation algorithms, Since Feb. 2023. Jan Ramon.

11.2.3 Juries

• Marc Tommasi member of the PhD jury of Victor Connes (Reviewer), Othmane Marfoq (Reviewer), Eduardo Brandao (President).
• Michaël Perrot: member of the PhD jury of Ganesh del Grosso (Examiner)
• Jan Ramon: member of the PhD jury of Angelo Saadeh, Les Applications du Calcul Multipartite Sécurisé en Apprentissage Automatique (Examiner)
• Jan Ramon: member of the PhD jury of Arnaud Griver-Sibert, Combining differential privacy and homomorphic encryption for privacy-preserving collaborative machine learning (Reviewer)
• Jan Ramon member of the PhD jury of Juan Alvarado, Modeling relational structures via constraint graphons (dept of computer science KULeuven). (Co-director)
• Aurélien Bellet: member of the PhD jury of Marina Constantini (Examiner)

11.3.1 Articles and contents

• Michaël Perrot: Article on the Inria Website on “AI decentralized: how to ensure more fairness and privacy?”

11.3.2 Interventions

• Michaël Perrot: Pint of Science “IA d'la joie” on fairness in artificial intelligence, Médiathèque de la Madeleine, France.
• Michaël Perrot: Utopiades organized by the Alter/Echos association on fairness in artificial intelligence, Remotely.

International journals

• 13 articleF.Florent Capelli, N.Nicolas Crosetti, J.Joachim Niehren and J.Jan Ramon. Linear Programs with Conjunctive Database Queries.Logical Methods in Computer ScienceJanuary 2024HALback to textback to text
• 14 articleG.Gaurav Maheshwari and M.Michaël Perrot. FairGrad: Fairness Aware Gradient Descent.Transactions on Machine Learning Research JournalAugust 2023HALback to text
• 15 articleA.Aishik Mandal, M.Michaël Perrot and D.Debarghya Ghoshdastidar. A Revenue Function for Comparison-Based Hierarchical Clustering.Transactions on Machine Learning Research JournalApril 2023HALback to text
• 16 articleC.César Sabater, F.Florian Hahn, A.Andreas Peter and J.Jan Ramon. Private Sampling with Identifiable Cheaters.Proceedings on Privacy Enhancing Technologies202322023HALback to text
• 17 articleL.Leander Schietgat, B.Bertrand Cuissart, K. D.Kurt De Grave, K.Kyriakos Efthymiadis, R.Ronan Bureau, B.Bruno Crémilleux, J.Jan Ramon and A.Alban Lepailleur. Automated detection of toxicophores and prediction of mutagenicity using PMCSFG algorithm.Molecular Informatics2023HALDOIback to text
• 18 articleA. S.Ali Shahin Shamsabadi, B. M.Brij Mohan Lal Srivastava, A.Aurélien Bellet, N.Nathalie Vauquier, E.Emmanuel Vincent, M.Mohamed Maouche, M.Marc Tommasi and N.Nicolas Papernot. Differentially private speaker anonymization.Proceedings on Privacy Enhancing Technologies20231January 2023HALDOIback to text

Invited conferences

• 19 inproceedingsA.Alberto Pedrouzo-Ulloa, J.Jan Ramon, P.Patrick Duflot, F.Fernando Pérez-González, S.Siyanna Lilova, Z.Zakaria Chihani, N.Nicola Gentili, P.Paola Ulivi, M. A.Mohammad Ashadul Hoque, T.Twaha Mukammel, Z.Zeev Pritzker, A.Augustin Lemesle, J.Jaime Loureiro-Acuña, X.Xavier Martı́nez and G.Gonzalo Jiménez-Balsa. Introducing the TRUMPET project: TRUstworthy Multi-site Privacy Enhancing Technologies.IEEE CSR 2P-DPA workshop - Workshop on Privacy-Preserving Data Processing and AnalysisVenice, ItalyJuly 2023HALback to text

International peer-reviewed conferences

• 20 inproceedingsB. L.Batiste Le Bars, A.Aurélien Bellet, M.Marc Tommasi, E.Erick Lavoie and A.-M.Anne-Marie Kermarrec. Refined Convergence and Topology Learning for Decentralized SGD with Heterogeneous Data.Proceedings of The 26th International Conference on Artificial Intelligence and Statistics (AISTATS 2023)Valencia, Spain, Spain2023HALback to text
• 21 inproceedingsE.Edwige Cyffers, A.Aurélien Bellet and D.Debabrota Basu. From Noisy Fixed-Point Iterations to Private ADMM for Centralized and Federated Learning.Proceedings of the 40th International Conference on Machine Learning (ICML)Honolulu, United StatesJuly 2023HALback to text
• 22 inproceedingsP.Pierre Humbert, B.Batiste Le Bars, A.Aurélien Bellet and S.Sylvain Arlot. One-Shot Federated Conformal Prediction.ICML 2023 - 40th International Conference on Machine LearningProceedings of the 40th International Conference on Machine Learning (ICML)Honolulu (Hawai), United StatesJuly 2023HALback to text
• 23 inproceedingsB.Bastien Liétard, M.Mikaela Keller and P.Pascal Denis. A Tale of Two Laws of Semantic Change: Predicting Synonym Changes with Distributional Semantic Models.Proceedings of the 12th Joint Conference on Lexical and Computational SemanticsThe 12th Joint Conference on Lexical and Computational SemanticsToronto, CanadaJune 2023HALback to text
• 24 inproceedingsG.Gaurav Maheshwari, A.Aurélien Bellet, P.Pascal Denis and M.Mikaela Keller. Fair Without Leveling Down: A New Intersectional Fairness Definition.EMNLP 2023 - The 2023 Conference on Empirical Methods in Natural Language ProcessingSingapore (SG), SingaporeDecember 2023HALback to text
• 25 inproceedingsP.Paul Mangold, A.Aurélien Bellet, J.Joseph Salmon and M.Marc Tommasi. High-Dimensional Private Empirical Risk Minimization by Greedy Coordinate Descent.International Conference on Artificial Intelligence and StatisticsAISTATS 2023 - International Conference on Artificial Intelligence and StatisticsValencia, SpainApril 2023HALback to text
• 26 inproceedingsP.Paul Mangold, M.Michaël Perrot, A.Aurélien Bellet and M.Marc Tommasi. Differential Privacy has Bounded Impact on Fairness in Classification.Proceedings of the 40th International Conference on Machine LearningInternational Conference on Machine Learning202Honolulu, United StatesJuly 2023HALback to text
• 27 inproceedingsC.Cennet Oguz, P.Pascal Denis, E.Emmanuel Vincent, S.Simon Ostermann and J.Josef van Genabith. Find-2-Find: Multitask Learning for Anaphora Resolution and Object Localization.2023 Conference on Empirical Methods in Natural Language ProcessingSingapore, Singapore2023HALback to text
• 28 inproceedingsJ.Joseph Renner, P.Pascal Denis, R.Rémi Gilleron and A.Angèle Brunellière. Exploring Category Structure with Contextual Language Models and Lexical Semantic Networks.Exploring Category Structure with Contextual Language Models and Lexical Semantic NetworksEACL 2023 - 17th Conference of the European Chapter of the Association for Computational LinguisticsDubrovnik, CroatiaMay 2023HALback to text
• 29 inproceedingsJ.Joseph Renner, P.Pascal Denis and R.Rémi Gilleron. WordNet Is All You Need: A Surprisingly Effective Unsupervised Method for Graded Lexical Entailment.Findings of the Association for Computational Linguistics: EMNLP 2023Singapore, France2023HALback to text
• 30 inproceedingsS.Sina Sajadmanesh, A. S.Ali Shahin Shamsabadi, A.Aurélien Bellet and D.Daniel Gatica-Perez. GAP: Differentially Private Graph Neural Networks with Aggregation Perturbation.USENIX Security 2023 - 32nd USENIX Security SymposiumAnaheim, United StatesAugust 2023HALback to text
• 31 inproceedingsD.Damien Sileo and M.-F.Marie-Francine Moens. Probing neural language models for understanding of words of estimative probability.Proceedings of the 12th Joint Conference on Lexical and Computational Semantics (*SEM 2023)Toronto, FranceAssociation for Computational LinguisticsJuly 2023, 469-476HALDOIback to text

Scientific books

• 32 bookI.Ilse Depraetere, B.Bert Cappelle, M.Martin Hilpert, L. D.Ludovic De Cuypere, M.Mathieu Dehouck, P.Pascal Denis, S.Susanne Flach, N.Natalia Grabar, C.Cyril Grandin, T.Thierry Hamon, C.Clemens Hufeld, B.Benoît Leclercq and H.-J.Hans-Jörg Schmid. Models of Modals: From Pragmatics and Corpus Linguistics to Machine Learning.Topics in English Linguistics [TiEL]110De Gruyter Mouton2023HALback to text

Doctoral dissertations and habilitation theses

• 33 thesisP.Paul Mangold. Exploiting Problem Structure in Privacy-Preserving Optimization and Machine Learning.Université de LilleOctober 2023HALback to textback to text
• 34 thesisA.Arijus Pleska. Privacy-preserving Learning by Averaging in Collaborative Networks.Université de LilleJune 2023HALback to textback to text

Reports & preprints

• 35 reportL.Louise Bart, E. A.El Amine Bechorfa, A.Antoine Boutet, J.Jan Ramon and C.Carole Frindel. A Smartphone-based Architecture for Prolonged Monitoring of Gait.Insa Lyon; Inria LyonDecember 2023HALback to text
• 36 miscL.Louis Béthune, T.Thomas Masséna, T.Thibaut Boissin, C.Corentin Friedrich, F.Franck Mamalet, A.Aurélien Bellet, M.Mathieu Serrurier and D.David Vigouroux. DP-SGD Without Clipping: The Lipschitz Neural Network Way.May 2023HALback to text
• 37 miscS.Shayne Longpre, R.Robert Mahari, A.Anthony Chen, N.Naana Obeng-Marnu, D.Damien Sileo, W.William Brannon, N.Niklas Muennighoff, N.Nathan Khazam, J.Jad Kabbara, K.Kartik Perisetla, A.Alexis Wu, E.Enrico Shippole, K.Kurt Bollacker, T.Tongshuang Wu, L.Luis Villa, S.Sandy Pentland and S.Sara Hooker. The Data Provenance Initiative: A Large Scale Audit of Dataset Licensing & Attribution in AI.November 2023HALback to textback to text
• 38 miscC.Clément Pierquin, A.Aurélien Bellet, M.Marc Tommasi and M.Matthieu Boussard. Rényi Pufferfish Privacy: General Additive Noise Mechanisms and Privacy Amplification by Iteration via Shift Reduction Lemmas.December 2023HALback to text
• 39 miscD.Damien Sileo and A.Antoine Lernould. MindGames: Targeting Theory of Mind in Large Language Models with Dynamic Epistemic Modal Logic.May 2023HALback to text
• 40 miscD.Damien Sileo. tasksource: A Dataset Harmonization Framework for Streamlined NLP Multi-Task Learning and Evaluation.May 2023HALback to textback to text