6.1.1 OATs'inside

• Keywords:
  Android, Malware, Reverse engineering, Code analysis
• Scientific Description:
  OATs'inside is an analysis tool that handles native Android applications. The system uses a hybrid approach based on dynamic monitoring and trace-based symbolic execution to output control flow graphs for each method of the analyzed application.
• Functional Description:
  OATs’inside is an Android reverse engineering tool that try to handle some native based obfuscation techniques. This tool uses a hybrid approach based on dynamic monitoring and trace-based symbolic execution to output control flow graphs (CFGs) for each method of the analyzed application. These CFGs spare users the need to dive into low-level instructions, which are difficult to reverse engineer.
• News of the Year:
  The full source code have been published and documented. New experiments have been released in september 2022 that enhance the results for more complex applications.
• URL:
  https://­gitlab.­inria.­fr/­cidre-public/­oatinside
• Publication:
  hal-02877815
• Authors:
  Pierre Graux, Jean-François Lalande, Valerie Viet Triem Tong, Pierre Wilke
• Contact:
  Jean-François Lalande

6.1.2 MoM

• Name:
  Malware-O-Matic
• Keywords:
  Malware, Cybersecurity, Ransomware
• Functional Description:
  MoM is an automated platform for conducting dynamic malware scans running on Windows. MoM is a bare-metal, non-virtualized platform on which user activity is simulated.
• Release Contributions:
  Refactoring allowing greater flexibility in its deployment and use. Monitoring of experiments.
• URL:
  https://­lhs-pec.­inria.­fr/­hosting/
• Publication:
  hal-01405636
• Contact:
  Valerie Viet Triem Tong
• Partner:
  DGA-MI

6.1.3 DaViz

• Name:
  Dataset Vizulisation
• Keywords:
  Visualization, Android
• Scientific Description:
  With millions of Android malware samples available, researchers have a large amount of data to perform malware detection and classification, specially with the help of machine learning. Thus far, visualization tools focus on single samples or one-to-many comparison, but not a many-to-many approach. Daviz is a web frontend/backend that aids to compare and explore Android application datasets. With the aid of multiple chart types and a system of interactive sample filtering, users can get a better understanding of the datasets at hand.
• Functional Description:
  Daviz is a web frontend and backend for the interactive visualization of large scale dataset of Android applications.
• News of the Year:
  A first version is in production in the LHS.
• Contact:
  Jean-François Lalande
• Participants:
  Tomas Concepcion Miranda, Leopold Ouairy, Damien Gourbeyre

Data and model to help the understanding of advanced attack campaigns.

Advanced Persistent Threat (APT) attacks are surgically targeted attacks led by advanced attackers, who constantly adapt their Tactics, Techniques, and Procedures (TTP).Identifying patterns in the modus operandi of attackers is an essential requirement in the study of Advanced Persistent Threats. Our community is hampered both by the lack of formalism helping to precisely describe these attacks and by the lack of accurate, relevant, and representative datasets of these current threats. In 2 we propose a formal model of an attacker’s tactical progression during the network propagation phase. This formalization allows to describe the PWNJUTSU experiment unequivocally. In this experiment, 22 Red Teamers attacked a vulnerable infrastructure to compromise machines and steal secret flags. we had three distinct goals in this experiment. A first goal was to observe how professionals attackers progress in such an infrastructure. A second goal was to build a dataset of various attacks targeting the same infrastructure. A last goal was to test the relevance of our model of a whole attack campaign. The resulting dataset is available online on here. It contains all participants’ event logs (system and network), including the reference environment’s event logs. A search engine is also provided to peek into the dataset.

Assessing the opportunity of combining state-of-the-art Android malware detectors

Research on Android malware detection based on Machine learning has been prolific in recent years. In 5, in collaboration with the TruX team at the University of Luxembourg we detail a large-scale evaluation of four state-of-the-art approaches that their achieved performance fluctuates when applied to different datasets. Combining existing approaches appears as an appealing method to stabilise performance. We therefore proceed to empirically investigate the effect of such combinations on the overall detection performance. In our study, we evaluated 22 methods to combine feature sets or predictions from the state-of-the-art approaches. Our results showed that no method has significantly enhanced the detection performance reported by the state-of-the-art malware detectors. Nevertheless, the performance achieved is on par with the best individual classifiers for all settings. Overall, we conduct extensive experiments on the opportunity to combine state-of-the-art detectors. Our main conclusion is that combining state-of-the-art malware detectors leads to a stabilization of the detection performance, and a research agenda on how they should be combined effectively is required to boost malware detection. All artifacts of this study form a dataset of $\tilde{0}.5$ million apks and all extracted features. These artifacts are made available for replicability here.

Quality of Android malware datasets.

Android security has received a lot of attention over the last decade, especially when working on malware analysis. The CIDRE team have also a long past work on malware analysis. When presenting our contributions in conferences or discussing with other researchers of the community we had the feelings that the datasets that are used in the literature were of low quality: if the datasets are outdated or not representative of the studied population, the conclusions may be flawed. In 4, we have investigated the irregularities of datasets used in experiments, questioning the validity of the performances reported in the literature. We have developed a new method for debiasing datasets. With this method we have rebuilt new dataset, more up-to-date and with less bias than in the literature. This dataset has been released on the artifact repository of IEEE, as an independent contribution for ensuring the reproducibilty of our experiments and to serve as new start point for future works. In particular, this dataset can be used by other reasearchers contributing in Android malware detection or classification with machine learning algorithms.

Imbalanced Classification with TPG Genetic Programming: Impact of Problem Imbalance and Selection Mechanisms

19

Model Stealing Attacks Against Inductive Graph Neural Networks

18

Errors in the CICIDS2017 dataset and the significant differences in detection performances it makes.

25

Among the difficulties encountered in building datasets to evaluate intrusion detection tools, a tricky part is the process of labelling the events into malicious and benign classes. The labelling correctness is paramount for the quality of the evaluation of intrusion detection systems but is often considered as the ground truth by practitioners and is rarely verified. Another difficulty lies in the correct capture of the network packets. If it is not the case, the characteristics of the network flows generated from the capture could be modified and lead to false results.

In this paper, we present several flaws we identified in the labelling of the CICIDS2017 dataset and in the traffic capture, such as packet misorder, packet duplication and attack that were performed but not correctly labelled. We also assess the impact of these different corrections on the evaluation of supervised intrusion detection approaches.

Detecting APT through graph anomaly detection.

22

Despite fruitful achievements made by unsupervised machine learning-based anomaly detection for network intrusion detection systems, they are still prone to the issue of high false alarm rates, and it is still difficult to reach very high recalls. In 2020, our team proposed (Leichtnam' PhD thesis) Sec2graph, an unsupervised approach applied to security objects graphs that exhibited interesting results on single-step attacks. The graph representation and the embedding allowed for better detection since it creates qualitative features.

In this paper, we present new experiments to assess the performances of this approach for detecting APT attacks. We achieve better detection performances than the original work's baseline detection methods on the DAPT2020 dataset.

Towards generic quality assessment of synthetic traffic for evaluating intrusion detection systems.

24

Network Intrusion Detection Systems (NIDSes) evaluation requires background traffic. However, real background traffic is hard to collect. We hence rely on synthetic traffic generated especially for this task. The quality of the generated traffic has to be evaluated according to some clearly defined criteria.

In this paper, we show how to adapt the quality assessment solutions proposed for different fields of data generation such as image or text generation to network traffic. We discuss the criteria that allow evaluation of the quality of a generated network traffic and propose functions to evaluate these criteria.

Cross-domain Alert Correlation methodology for Industrial Control Systems

Alert correlation is a set of techniques used to process alerts raised by various intrusion detection systems in order to a eliminate redundant alerts, reduce the number of false alerts, and reconstruct attack scenarios. In Industrial Control Systems, the presence of a physical process and the associated specific threats has led to the heterogeneity of alerts due to the development of multi-domain detection techniques. Some detection approaches rely solely on observations at the level of the cyber domain while other approaches monitor the physical process. The two approaches are complementary but the information carried by the two types of alerts are different. In 6, we combine the alerts from physical domain intrusion detection with more classical cyber-domain intrusion detection alerts. We develop an alert correlation approach using an alert enrichment that allows mapping physical domain alerts into the cyber domain. We also propose a specific alert selection for correlation that adapts to the state of the physical process by dynamically adjusting the size of the selected alert window. We publicly released all the datasets generated and used in our results.

Intrusion Detection through Lightweight Tangled Program Graphs

The fast improvement of Machine-Learning (ML) methods gives rise to new attacks in Information System (IS). Simultaneously, ML also creates new opportunities for network intrusion detection. Early network intrusion detection is a valuable asset for IS security, as it fosters early deployment of countermeasures and reduces the impact of attacks on system availability. In 8 we propose and study an anomaly-based Network Intrusion Detection System (NIDS) based on Tangled Program Graph (TPG) ML. Secure-GEGELATI learns how to detect intrusions from IS-produced traces and is optimised to fit the requirements of intrusion detection. The study evaluates the capacity of Secure-Gegelati to act as a continuously learning, real-time, and low energy NIDS when executed in an embedded network probe. We show that a TPG is capable of switching between training and inference phases, new training phases enriching the probe knowledge with limited degradation of previous intrusion detection capabilities. The Secure-GEGELATI software reaches 8x the energy efficiency of an optimised Random Forests (RF)-based Intrusion Detection System (IDS) on the same platform. It is capable of processing 13.2 k connections/seconds with a peak power of less than 3.3 Watts on an embedded platform, and is processing in real-time the CIC-IDS 2017 dataset while detecting $84\%$ of intrusions and raising less than $0.2\%$ of false alarms.

Anomaly-based Intrusion Detection using POSET

In the context of the European SPARTA project, we have proposed an anomaly-based intrusion detection tool where the normal behavior of an application is described using models based on automata and invariants. First, the solution was evaluated using a distributed application (XtreemFS, a distributed file system). In this case, the events observed on different machines form a partially ordered set (POSET), which is the ideal target of our solution. Then, the tool was adapted to analyze network traffics (both Pcap and CSV files). Here observed events are totally ordered. Results of these evaluations are summarized in the published SPARTA deliverables.

Exploring Federated Prediction of Security Events

Prior works in industrial intrusion detection practices follow the general approach to build AI-driven predicative models to learn from historical data. Such models characterize previous attack events and use this knowledge to predict future ones. These systems typically require collecting events from different industrial organizations and storing them in a centralized AI-as-a-Service (AIaaS) platform to train an AI-driven prediction model. However, these records often include privacy-sensitive metadata, including machine ID, event description, timestamp, system hardening actions taken, etc. Moreover, disclosing them can reveal sensitive information about security policies and security postures. Confidentiality concerns of industrial customers hence make it unrealistic to deploy such a centralized learning process of AI-based security event prediction methods. In 15, we investigate the feasibility of using privacy-friendly collaborative learning to benefit from participating organizations’ knowledge without requiring data disclosure. In particular, we turn to employ Federated Learning to train AI models collaboratively by first locally training AI models per participating organization and then aggregating the local model updates. This Federated AI-based intrusion detection method is tested using real-world network attack attack data stored in a distributed way by 20 different industrial participants. We analyze the detection performance of this Federated AI system in practices, including measuring the detection performances and communication / computational cost of the Federated AI system. Furthermore, we discuss how the distribution drift of the attack behaviors across different industrial users affect the accuracy of the Federated AI system. We demonstrate how to measure the contribution of different participating organizations in the system to the collectively trained intrusion detection model and evaluate the benefits of each organization gained from joining the federation network. Finally, we unveil that distributed data poisoning attacks against the Federated AI system are effective at undermining robustness and decreasing the attack detection precision by a significant amount. While the state-of-the-art countermeasures are claimed to be effective, we demonstrate that carefully tuned data poisoning attacks can easily bypass the defense measures and injects decision bias to the Federated AI system without triggering any alerts of the hygiene check. Besides, we show that data reconstruction attacks can be performed against the participating organizations of the Federated AI system by leaking the statistical profiles of their privately owned attack data used for AI model training. Efficient as the Federated AI system is, our study also raises a severe concern over the potential risk of system integrity and data privacy leaks inside this Federated AI system.

AdvCat: Domain-Agnostic Robustness Assessment for Cyber security-Critical Applications with Categorical Inputs

Machine Learning-as-a-Service systems (MLaaS) have been largely developed for cyber security-critical applications, such as detecting network intrusions and fake news campaigns. Despite effectiveness, their robustness against adversarial attacks is one of the key trust concerns for MLaaS deployment. Our work 16 are thus motivated to assess the adversarial robustness of the Machine Learning models residing at the core of these security-critical applications with categorical inputs. Previous research efforts on accessing model robustness against manipulation of categorical inputs are specific to use cases and heavily depend on domain knowledge, or require white-box access to the target ML model. Such limitations prevent the robustness assessment from being as a domain-agnostic service provided to various real-world applications. We propose a provably optimal yet computationally highly efficient adversarial robustness assessment protocol for a wide band of ML-driven cyber security-critical applications. We demonstrate the use of the domain-agnostic robustness assessment method with substantial experimental study on fake news detection and intrusion detection problems. ——

Identity Management Systems

allow users to prove characteristics about themselves to multiple service providers. IMS evolved from impractical, site-by-site authentication, to versatile, privacy enhancing Self Sovereign Identity (SSI) Frameworks. SSI frameworks often use Anonymous Credential schemes to provide user privacy, and more precisely unlinkability between uses of these credentials. However, these schemes imply the disclosure of the identity of the Issuer of a given credential to any service provider. This can lead to information leaks. In 3, we have proposed a new Anonymous Credential scheme that allows a user to hide the Issuer of a credential, while being able to convince the service providers that they can trust the credential, in the absence of a trusted setup. We proved this new scheme secure under the Computational Diffie Hellman assumption, and Decisional Diffie Hellman assumption, in the Random Oracle Model. We show that this scheme is efficient enough to be used with laptops, and to be integrated into SSI frameworks or any other IMS.

Boundaries Extending The Boundaries and Exploring The Limits Of Blockchain Compression

Blockchain technology aims to replace traditional banking systems and manage the world’s economic data. However, the long-term feasibility of blockchain technology is hindered by the inability of existing blockchain protocols to prune the consensus data leading to constantly growing storage and communication requirements. Kiayias et al. have proposed a blockchain protocol based on superblock Non-InteractiveProofs-of-Proof-of-Work (NIPoPoWs) as a mechanism to reduce the storage and communication complexity of blockchains to $O\left(polylog\right(n\left)\right)$. However, their protocol is only resilient to an adversary that may control strictly less than 1/3rd of the total computational power, which is a reduction from the security guaranteed by Bitcoin and other existing blockchain protocols that guarantee security against an adversary that may control strictly less than $1/2$ of the total computational power. We present an improvement to the Kiayias et al. proposal termed Gems-scheme, which is resilient against an adversary that may control less than $1/2$ of the total computational power while operating in $O\left(polylog\right(n\left)\right)$ storage and communication complexity. Additionally, we present a novel proof that establishes a lower bound of $O\left(log\right(n\left)\right)$ on the storage and communication complexity of any PoW-based blockchain protocol 29.

Stochastic analysis of rumor spreading with multiple pull operations

We propose and analyze a new asynchronous rumor spreading protocol to deliver a rumor to all the nodes of a large-scale distributed network 7. This spreading protocol relies on what we call a k-pull operation, with $k\ge 2$. Specifically a k-pull operation consists, for an uninformed node s, in contacting $k-1$ other nodes at random in the network, and if at least one of them knows the rumor, then node s learns it. We perform a thorough study of the total number $T_{k,n}$ of k-pull operations needed for all the n nodes to learn the rumor. We compute the expected value and the variance of $T_{k,n}$, together with their limiting values when n tends to infinity. We also analyze the limiting distribution of $(T_{k,n}-E\left(T_{k,n}\right))/2$ and prove that it has a double exponential distribution when n tends to infinity. Finally, we show that when $k>2$, our new protocol requires less operations than the traditional 2-push-pull and 2-push protocols by using stochastic dominance arguments. All these results generalize the standard case $k=2$.

Sycomore$++$

The arrival of Bitcoin drove the shift to decentralized ecosystems through the exchange of transactions without intermediary. However, one of the main challenges that need to face permissionless blockchains are scalability and security. Sycomore$++$ is a permissionless graph-based distributed ledger whose main feature is to dynamically self-adapt the number of created blocks to the current number of submitted transactions 9, 13. We have conducted a performance evaluation of Sycomore$++$ and have compared it with the ones of Bitcoin and Sycomore, a graph-based distributed ledger. Our evaluation relies on agentbased simulations to evaluate the capability of these distributed ledgers to address the aforementioned challenges, within different execution contexts. One of the main lessons drawn from these intensive simulations is the capability of Sycomore$++$ to drastically reduce transaction confirmation time with respect to the other two ledgers, to quickly react to any sudden variation of the transaction submission rate, to minimize the computational power waste w.r.t. PoW-based permissionless distributed ledgers, and to surpass Bitcoin in terms of resilience to a network adversarial environment. We also study resilience to adversarial miners that want to endanger the quality of the graph showing that, as in Bitcoin, the adversary is limited by its computational power12.

Towards understanding the robustness against evasion attack on categorical inputs

Characterizing and assessing the adversarial vulnerability of classification models with categorical input has been a practically important, while rarely explored research problem. Categorical data exist widely in cyber security applications. For example, dynamic and static features of malware samples contain many categorical run-time and static signatures attributes, such as the types of API calls, files / processes visited or created and generated network traffics. Adversarial attacks against ML-based security incident detection systems raise a severe concern over the trustworthiness of ML-based detection. Therefore, exploring the origin of adversarial risk over categorical data can deepen our understanding about the feasibility of adversarial threats against real-world cyber security applications.

Our work 10 echoes the challenge by first unveiling the impact factors of adversarial vulnerability of classification models with categorical data based on an information-theoretic adversarial risk analysis about the targeted classifier. Though certifying the robustness of such classification models is intrinsically an NP-hard combinatorial problem, our study shows that the robustness certification can be solved via an efficient greedy exploration of the discrete attack space for any measurable classifiers with a mild smoothness constraint. Our proposed robustness certification framework is instantiated with deep neural network models applied on real-world safety-critic data sources. Our empirical observations confirm the impact of the key adversarial risk factors with categorical input.

Finding MNEMON: Reviving Memories of Node Embeddings

In this work 17, we focus on exploring the potential data privacy leaks while deploying graph-based machine learning over the privately owned data, such as molecules invented to develop new medicines and social networks. Unveiling the privacy information about the training data directly violate the intellectual property policy of any entity using the Machine Learning service.

Previous security research efforts orbiting around graphs have been exclusively focusing on either (de-)anonymizing the graphs or understanding the security and privacy issues of graph neural networks. Little attention has been paid to understand the privacy risks of integrating the output from graph embedding models (e.g., node embeddings) with complex downstream machine learning pipelines. In this paper, we fill this gap and propose a novel model-agnostic graph recovery attack that exploits the implicit graph structural information preserved in the embeddings of graph nodes. We show that an adversary can recover edges with decent accuracy by only gaining access to the node embedding matrix of the original graph without interactions with the node embedding models. We demonstrate the effectiveness and applicability of our graph recovery attack through extensive experiments.

RT-DFI: Optimizing Data-Flow Integrity for Real-Time Systems

The emergence of Real-Time Systems with increased connections to their environment has led to a greater demand in security for these systems. Memory corruption attacks, which modify the memory to trigger unexpected executions, are a significant threat against applications written in low-level languages. Data-Flow Integrity (DFI) is a protection that verifies that only a trusted source has written any loaded data. The overhead of such a security mechanism remains a major issue that limits its adoption. In 11, we presents RT-DFI, a new approach that optimizes the Data-Flow Integrity to reduce its overhead on the Worst-Case Execution Time. We model the number and order of the checks and use an Integer Linear Programming solver to optimize the protection on the Worst-Case Execution Path. Our approach protects the program against many memory-corruption attacks, including Return-Oriented Programming and Data-Only attacks. Moreover, our experimental results show that our optimization reduces the overhead by 7% on average compared to a state-of-the-art implementation.

9.1.1 Visits of international scientists

9.1.2 Visits to international teams

9.1.3 H2020 projects

PEPR CyberSecurity project: DefMal (2022-2028)

Participants: Kevin Allix, Pierre-François Gimenez, Yufei Han, Jean-François Lalande, Valérie Viet Triem Tong.

PEPR Cybersecurity project: SecureEval(2022-2028)

Participants: Guillaume Hiet, Frédéric Tronel, Pierre Wilke.

PEPR Cybersecurity project: SuperviZ (2022-2028)

Participants: Pierre-François Gimenez, Gilles Guette, Yufei Han, Ludovic Mé.

ANR Project: Byblos (2021-2025).

Participants: Emmanuelle Anceaume.

ANR Project: BC4SSi (2023-2027)

Participants: Emmanuelle Anceaume.

CominLabs project: Priceless (2021-2025)

Participants: Emmanuelle Anceaume.

ANR Project: TrustGW (2021-2025).

Participants: Guillaume Hiet, Frédéric Tronel, Pierre Wilke.

PHC AURORA Project: SECUTRACE (2020-2022).

Participants: Guillaume Hiet.

CominLabs project: SCRATCHS (2021-2024)

Participants: Pierre Wilke, Guillaume Hiet.

10.1.1 Scientific events: organisation

10.1.2 Scientific events: selection

10.1.3 Journal

10.1.4 Invited talks

• Ludovic Mé was panelist for a round table organized by BNP-Paribas and dedicated to the interaction between cyber security and AI (April 2022, 14th).
• Emmanuelle Anceaume was invited to give a talk at BRAINS 2022
• Pierre-François Gimenez was invited to give a talk at SUPSEC
• Guillaume Hiet and Pierre Wilke were invited to give talks at the COEMS Forsterk Seminar 2022

10.1.5 Scientific expertise

Jean-Francois Lalande was a reviewer for the phD grants “RIN Doctorants” of Normandie University and for the call for projects IRGA 2022 of Grenoble Alpes University.

Ludovic Mé serves:

• the Scientific Council of the LIRIMA (Laboratoire International de Recherche en Informatique et Mathématiques Appliquées) ;
• the Expert Council of the DST  (Digital Science and Technology Network) ;
• the “Bureau du GT sécurité des systèmes logiciels” of the GDR “sécurité” ;
• the expert group for the evaluation of French research entities (UMRs and EAs) relatively to the protection of scientific and technological properties (PPST).

Guillaume Hiet served as a reviewer for an ANR project.

Since October 2022, Guillaume Hiet is the co-chair of the Systems, Software and Network Security working group of the GDR Sécurité Informatique.

10.1.6 Research administration

Ludovic Mé is deputy scientific director of Inria, in charge of the cyber security area.

The team participated to several recruitment committees:

• Ludovic Mé was member of the recruitment committees for a Professor position and for an Assistant Professor position at Télécom-Paris.
• Valérie Viet Triem Tong was member of the recruitment committees for a Professor position at University of Lorraine.
• Valérie Viet Triem Tong was member of the recruitment committees for a Professor position at IMT.
• Valérie Viet Triem Tong was member of the recruitment committees for an Assistant Professor position at CentraleSupélec.
• Valérie Viet Triem Tong was member of the recruitment committees for a Professor position at CentraleSupélec.
• Emmanuelle Anceaume was member of the recruitment committees for a Professor position at ESIR.
• Emmanuelle Anceaume was member of the recruitment committees for a Maitre de Conférence position at ISTIC.
• Jean-François Lalande was member of the recruitment committees for an Assistant Professor position at CentraleSupélec.
• Jean-François Lalande was member of the recruitment committees for a Professor position at CentraleSupélec.
• Guillaume Hiet was member of the recruitment committees for an Assistant Professor position at CentraleSupélec.

10.2.1 Teaching

Several team members are involved in initial and continuing education in CentraleSupélec, a french institute of research and higher education in engineering and science, ESIR (Ecole Supérieure d'Ingénieur de Rennes) the graduate engineering school of the University of Rennes 1.

In these institutions,

• Christophe Bidan is the head of the Rennes campus of CentraleSupélec;
• Gilles Guette is the director of corporate relations at ESIR;
• Jean-François Lalande is responsible of the major program dedicated to information systems security and the special track Infosec of CentraleSupélec;
• Frédéric Tronel and Valérie Viet Triem Tong share the responsability of the mastère spécialisé (post-graduate specialization degree) in Cybersecurity. This education was awarded best French master degree in the category “Master Cybersecurity masters and Security of systems” in the Eduniversal master ranking 2021.

The teaching duties are summed up in table 1.