5.1.1 Blare

• Name:
  Blare, an information flow monitor
• Keywords:
  Cybersecurity, Intrusion Detection Systems (IDS), Data Leakage Protection
• Scientific Description:
  Blare implements our approach of illegal information flow detection for a single node (Android and Linux kernel, JVM) and a set of nodes (monitoring of flows between linux machines).
• Functional Description:
  Blare is an information flow monitor that opérates at the OS level. Blare relies on tainting techniques to monitor information flow between files, processes, sockets and memory pages. Blare allows to identify how a (malicious) application contaminates the OS.
• News of the Year:
  In 2021, we have worked on porting Blare to new recent versions of the Linux kernel. In particular, we started to work on kernels 4.x that are used on mobile phone in order to be able to build AndroBlare, a modified version of Android based on Android 9 or 10.
• URL:
  http://­www.­blare-ids.­org
• Publications:
  hal-01535949, hal-01535862, hal-00268408, hal-00356441, hal-00356484, hal-00420117, hal-00875211, hal-00840338, hal-00909400, hal-00862468, hal-00736045, hal-00736034, hal-00647116, hal-00647170, hal-00736045
• Contact:
  Valerie Viet Triem Tong
• Participants:
  Pierre Wilke, Valerie Viet Triem Tong, Frédéric Tronel, Alexandre Sanchez, Jean-François Lalande, Laurent Georget, Guillaume Hiet, Christophe Hauser, Radoniaina Andriatsimandefitra Ratsisahan
• Partner:
  CentraleSupélec

5.1.2 GroddDroid

• Name:
  GroddDroid
• Keywords:
  Android, Detection, Malware
• Scientific Description:
  GroddDroid automates the dynamic analysis of a malware. When a piece of suspicious code is detected, GroddDroid interacts with the user interface and eventually forces the execution of the identified code. Using Blare (Information Flow Monitor), GroddDroid monitors how an execution contaminates the operating system. The output of GroddDroid can be visualized in an web browser. GroddDroid is used by the Kharon software.
• Functional Description:
  GroddDroid 1 - locates suspicious code in Android application 2 - computes execution paths towards suspicious code 3 - forces executions of suspicious code 4 - automates the execution of a malware or a regular Android application
• Release Contributions:
  - Increase of the speed of the static analysis - New language for targeting the suspicious code - New strategy for discovering the GUI using the Manifest and the layout files - Visualization of the bytecode events on a new timeline - Integration tests have been created
• News of the Year:
  In 2021, we have created integration tests, corrected a large number of bugs and improved the visualization of the bytecode. We continue to migrate AndroBlare on more recent versions of Android.
• URL:
  http://­kharon.­gforge.­inria.­fr/­grodddroid.­html
• Publications:
  hal-01311917, hal-01201743, hal-01584989, hal-01535678
• Authors:
  Mourad Leslous, Adrien Abraham, Pierre Graux, Jean François Lalande, Valerie Viet Triem Tong, Pierre Wilke
• Contact:
  Jean-François Lalande
• Partners:
  CentraleSupélec, Insa Centre Val-de-Loire

5.1.3 OATs'inside

• Keywords:
  Android, Malware, Reverse engineering, Code analysis
• Functional Description:
  OATs’inside is an Android reverse engineering tool that handles all native 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 code has been released using an open source license.
• Publication:
  hal-02877815
• Authors:
  Pierre Graux, Jean-François Lalande, Valerie Viet Triem Tong, Pierre Wilke
• Contact:
  Pierre Graux

5.1.4 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

5.1.5 TEMPO

• Name:
  The Evasive Malware PlatfOrm
• Keyword:
  Malware
• Functional Description:
  This platform enables reseachers to label Windows malware and download a list of evasive samples.
• URL:
  https://­tempo.­irisa.­fr
• Contact:
  Cédric Herzog

5.1.6 GUI-Mimic

• Keywords:
  Malware, Software testing
• Functional Description:
  GUI-Mimic is a software that automates the use of graphical software through short recorded sequences. Indeed, there is no tool that allows to easily generate varied and randomized sequences that stimulate many components of the same software. GUI-Mimic was developed for malware analysis.
• URL:
  https://­gitlab.­inria.­fr/­vraulin/­GUI-Mimic
• Publication:
  hal-03449827
• Contact:
  Vincent Raulin

Privacy Stealing Attacks Against Inductive Graph Neural Networks.

Graph Neural Networks (GNNs) are a variant of machine learning techniques that can be applied to exploit topological information contained by graph-structured data. In 1, we unveil how these graph learning models are prone to the privacy-motivated attacks. Such privacy-stealing efforts can clone the decision behaviors of the target graph learning models via merely querying the output of the models. Such privacy stealing efforts can facilitate downstream data privacy reserve engineering attacks as a stepping stone, like membership inference and model inversion attacks.

Our study defines a threat model to characterize an adversary’s background knowledge along two dimensions. We establishes six different attack scenarios based on this threat model. Finally we extensively evaluate our proposition on a six benchmark graph datasets and demonstrate the efficacy of these attacks. Interestingly, we unveil that the model privacy stealing can be achived by simply querying the classification output of the target graph learning models. It indicates that it is paramount to consider model/data privacy protection in applying Machine Learning-as-a-Service in real-world privacy-critical applications.

Visualization and monitoring

Romain Brisse started his PhD thesis in the Malizen company, which is a start-up that originated in the CIDRE team. Malizen develops the ZeroKit tool that aims to investigate complex attacks by using a visual tool offering interactive views of logs (system, network). During his first year, Romain proposed a new recommender system that helps the analyst to choose among proposals for his next step of investigation 9. These recommendations increase the efficiency of the investigator. These recommendations are computed using the knowledge extracted from the MITRE ATT&CK (real world adversary knowledge) and scored by algorithms that manipulate the investigated logs. Experiments have been conducted with experts in log analysis on the dataset TC3 from the DARPA in order to show the usefulness of the recommendations.

Effect of electromagnetic fault injection on micro-architecture

In 7, we demonstrate that complex CPUs are vulnerable to fault injection attacks. In particular, the memory hierarchy and the MMU can be changed, which creates a mismatch between hardware behavior and software expectations. In this work, we highlight several new fault effects at the micro-architectural level, in particular in the memory hierarchy: On the L1 instruction cache, on the memory management, on the L2 cache.

Low-Cost Evaluation Platform for Multifault Injection

Fault injection using laser beams or electromagnetic injection (EMI) are used in the literature to hijack control flow or to change a register value. Experiments in this domain requires expensive experiments due to the cost of EMI benches or laser probes. In 10, we propose TRAITOR, a low-cost evaluation platform to precisely inject numerous faults by clock glitch. With a low-cost FPGA, we produce our own faulty clock signal to replace the one of the targeted board. The platform makes the injection more controllable and easily reproducible. It provides the ability to inject consecutive faults and to realize complex attacks inducing several bursts of faults.

Vulnerability detection

The results obtained during the PhD thesis of Pierre Graux led to investigate attacks that use both the Java bytecode and the native code to flaw some sensitive data in an Android application. Such attack would be possible if a developer misuse the "transient" keyword on a sensitive field that would be serializable. We proposed an automatic detection of such vulnerability by combining static analysis and symbolic execution 12. Our approach was applied on the Telegram source code and we found non-exploitable flows hidden within Telegram's code base.

Stalkerware detection 

Stalkerware enables adversaries to conduct surveillance and remote control on a targeted person’s device. The Android devices are a particularly fertile ground for stalkerware, most of which spy on a single communication channel, sensor, or category of private data, though 27% of stalkerware surveil multiple private data sources. In 18, we propose DOSMELT, a stalkerware detection system with nuanced warnings that precisely characterize the types of surveillance functionalities conducted by Android stalkerware. The goal of DOSMELT is to provide in-time alerts to victims of stalkerwares. They can then take appropriate mitigating actions. Furthermore, we introduce active learning techniques into the design of DOSMELT, to promote effective data-driven security analysis even with few manually annotated apps.

Threat Hunting

Threat hunting is the process of searching through a compromised network to isolate an active attacker. The efficiency of this process depends on the defender's ability to effectively identify the traces left by the attacker in the network. In 3, we formalize both defensive processes and the attacker’s offensive approaches, allowing for confronting their respective perceptions during the same attack campaign. The attacker’s perception of the campaign is built from the execution of his attack procedures, his exposed resources and the compromised components. The defender’s perception of the attack is built from the collected traces on the targeted information system. This model leads to the construction of two persistent graphs on a common set of objects and components allowing for (1) an omniscient actor to compare, for both defender and attacker, the gap in knowledge and perceptions; (2) the attacker to become aware of the traces left on the targeted network; (3) the defender to improve the quality of Threat Hunting by identifying false-positives and adapting logging policy to be oriented for investigations.

Preventing hardware timing information leakage.

Since Spectre and Meltdown attacks were published in 2018, numerous attacks target the whole microarchitecture to extract information from timing variations. In 11, we propose and implement a process to build cores without microarchitectural timing leakage. We outline new generic design rules based on first principles to prevent timing information leakage, and build whole cores immune to them.

Injecting protections at runtime using Dynamic Binary Instrumentation.

Memory corruption attacks have been a significant issue in software security for over two decades and are still one of the most dangerous and widespread types of attacks nowadays. In 13, we propose an approach to protect applications against memory corruption attacks at runtime. We developed DAMAS, a framework capable of deploying such protections on running applications. Our solution does not require recompiling the application and uses runtime information to optimize the protections during the process execution. We implemented a Control-Data Isolation protection using our framework and evaluated its impact on performance.

Blockchain in adversarial environments.

We are pursuing our efforts dedicated to the theoretical aspects of blockchains in adversarial environments. This work, which started last year, aims at formally characterizing what a user can expect from blockchains in presence of adversarial environments. We argue that such an expectation can be fully characterized by the notion of « finality ». The notion of finality refers to the time when it becomes impossible to remove a block that has previously been appended to the blockchain. Blockchain finality can be deterministic or probabilistic, immediate or eventual. Informally, immediate finality guarantees, as its name suggests, that when a block is appended to a local copy, it is immediately finalized and thus will never be revoked in the future. Most of the permissionned blockchains satisfy the deterministic form of immediate consistency, including Red Belly blockchain and Hyperledger Fabric blockchain, while the probabilistic form of immediate consistency is typically achieved by permissionless pure proof-of-stake blockchains such as Algorand, or StateCube. Designing blockchains with immediate finality favors consistency against availability in presence of transient partitions of the system. It leverages the properties of Consensus (i.e a decision value is unique and agreed by everyone), at the cost of synchronization constraints.

On the other hand eventual finality ensures that all local copies of the blockchain share a common increasing prefix, and thus finality of their blocks increases as more blocks are appended to the blockchain. The majority of cryptoassets blockchains, with Bitcoin and Ethereum as celebrated examples, guarantee eventual finality with some probability: blocks may be revoked after being appended to the blockchain, yet with decreasing probability as they sink deeper into the chain.

While probabilistic eventual finality has been widely studied in the context of Bitcoin, some few studies have started to lay the foundations of the computation power of blockchains with deterministic eventual finality consistency.

We reinvestigate the definition of (deterministic) eventual prefix consistency introduced in 20 to fit both the context in which an infinite number of blocks are appended to the blockchain. We introduce the notion of bounded displacement, which informally says that the number of blocks that can be revoked from the current blockchain is bounded. Specifically, we show that known bounded displacement eventual prefix consistency is equivalent to Consensus, that unknown bounded displacement eventual prefix consistency is equivalent to eventual strong prefix consistency, and eventual strong prefix consistency is stronger than Eventual Consensus. We provide an algorithm that guarantees eventual prefix consistency in an asynchronous environment with an unbounded number of Byzantine processes. We also show that it is impossible to build a blockchain that guarantees eventual prefix consistency based on the longest chain rule. Finally, we discuss impossibilities and possibilities of unbounded displacement eventual prefix consistency. In particular, we propose an algorithm that solves unknown bounded displacement eventual prefix in an eventually synchronous environment in presence of less than a majority of Byzantine processes. To summarize, we close the gap between eventual prefix and strong prefix consistencies. More details appear in 8.

Uniform node sampling in adversarial environments.

This work considers the problem of achieving uniform node sampling in large scale systems in presence of Byzantine nodes. The uniform node sampling service offers to applications using it a single simple primitive that returns, upon invocation, the identifier of a random node that belongs to the system. We first propose an omniscient strategy that processes on the fly an unbounded and arbitrarily biased input stream made of node identifiers exchanged within the system, and outputs a stream that preserves the uniformity property. Informally, uniformity states that any node in the system should have the same probability to appear in the sample of any correct node of the system. We show through a Markov chain analysis that this property holds despite any arbitrary bias introduced by the adversary. We then propose a strategy based on a sketch data structure that is capable of approximating the omniscient strategy without requiring any prior knowledge on the composition of the input stream. We show through both theoretical analysis and extensive simulations that this "knowledge-free" strategy accurately approximates the omniscient one. We evaluate the resilience of the knowledge-free strategy by studying two representative attacks (flooding and targeted attacks). We quantify the minimum number of identifiers that Byzantine nodes must insert in the input stream to prevent uniformity. Finally, we propose a new construction that processes each input stream with sketches put in series that allows to both increase the accuracy of a single sketch and decrease the time to converge to a uniform output stream. To our knowledge, such a work has never been proposed before. More details appear in 2.

Stochastic analysis of rumor spreading.

The long term objective of this work is a deep analysis of rumor spreading in large-scale and open networks in presence of adversarial behaviors. So far, we have proposed and analyzed a new asynchronous rumor spreading protocol to deliver a rumor to all the nodes of a large-scale distributed network. This spreading protocol relies on what we call a k-pull operation for any value of k. 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 have performed a thorough study of the total number of k-pull operations needed for all the n nodes of the system to learn the rumor, for any value of n. All these results appear in 6, 14. The second step of this work will be to take into account the multiplicity of the recipients of the pull operations to design Byzantine-tolerant pull operations.

Stochastic analysis of algorithms for collecting longitudinal data.

This work proposes and analyses the performance and the vulnerability to attacks of three algorithms for collecting longitudinal data in a large scale system. A monitoring device is in charge of continuously collecting measurements from end-devices. The communication graph is connected but not necessarily complete. For scalability reasons, at each collect, a single end-device is randomly selected among all the end-devices to send the content of its local buffer of data to the monitoring device. Once sent, the end device resets its buffer, and resumes its measurement process. Two of the three algorithms are randomized algorithms while the third one is deterministic. The difference between the randomized algorithms stems from the random choice policy: in the first algorithm, choice is uniform while in the second one the random choice is weighted by the current amount of measurements at end-devices. The third algorithm is deterministic. End-devices are successively chosen in a round robin way. We study the transient and stationary maximum load distribution at end-devices when collects are made using the first and third algorithm, and by providing bounds via a coupling argument when the second algorithm is used. While the third algorithm provides the best performance, it is highly vulnerable to attacks. Details of this line of research is described in 15.

Safe artificial intelligence for trust-critical Machine-Learning-as-a-Service system.

With dramatic advance of Deep Learning (DL) techniques during the past decade, various DL algorithms have achieved impressive human-level classification and notable success in domains from computer vision to natural language processing. Increasingly deployed in various real-world trust-critical systems, safety of DL techniques becomes a leading research priority and is as important as the focus on accuracy, speed, and scalability. The stakeholders of DL services often have the safety concern and ask how they can trust the DL-enabled systems in the cost-sensitive decision-making process. In this proposal, our study is concretized to three trust-critical applications with categorical input data including 1) malware detection and intrusion detection 17, 2) graph-based social network analysis 1 and 3) automated medical diagnosis using electronic health records. Compared to continuous data, like images, establishing the safety guarantees on categorical inputs against various adversarial attacks raises a unique challenge: the adversarial attacks on categorical data are intrinsically NP-hard Mixed-Integer Non-Linear Programming problems. Our study thus aims to provide both theoretical and practical answers to assess and mitigate the adversarial vulnerability of Deep Learning techniques on categorical data, in order to deliver safety-guaranteed learning in trust-critical applications.

7 Bilateral contracts and grants with industry

8.1.1 FP7 & H2020 projects

8.1.2 Other european programs/initiatives

Participants: Jean-François Lalande, Valérie Viet Triem Tong, Pierre Wilke.

ANR Project: Byblos (2021-2025).

Participants: Emmanuelle Anceaume.

ANR Project: TrustGW (2021-2025).

Participants: Guillaume Hiet, Frédéric Tronel, Pierre Wilke.

PHC AURORA Project: SECUTRACE (2020-2021).

Participants: Guillaume Hiet.

CominLabs project: Priceless (2021-2025).

Participants: Emmanuelle Anceaume.

CominLabs project: SCRATCHS (2021-2024).

Participants: Guillaume Hiet, Frédéric Tronel, Pierre Wilke.

FUI project: SECEF (2020-2022).

Participants: Guillaume Hiet.

9.1.1 Scientific events: organisation

Seminar "Hands-on Machine Learning for Security", sponsored by the GDR Sécurité Informatique, organised by Pierre-François Gimenez and Yufei Han. This seminar was an in-person event and broadcasted on Youtube (link).

9.1.2 Scientific events: selection

9.1.3 Journal

9.1.4 Invited talks

Guillaume Hiet gave an invited talk at the (COEMS Forsterk Seminar) on Hardblare – heterogeneous solution for embedded software security.

Ludovic Mé was panelist for round tables dedicated to the interaction between cyber security and AI, in the context of:

• the conference CAID 2021 (Conference on Artificial Intelligence for Defense) ;
• the Atos Cyber Day.

9.1.5 Scientific expertise

Jean-Francois Lalande was a reviewer for the phD grants in IA for the "HAISCoDe" program of Normandie University.

Ludovic Mé serves:

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

9.1.6 Research administration

Emmanuelle Anceaume was member of a recruitment committee for a Maitre de conférence position at the University of Rennes 1.

Guillaume Hiet and Valérie Viet Triem Tong were members of a recruitment committee for a Maitre de conférence position at CentraleSupelec.

Ludovic Mé was member of a recruitment committee for a Professor position at the University of Rennes 1, and for a Starting Reseach Position at Inria.

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

Valérie Viet Triem Tong was member of a recruitment committee for a Maitre de conférence position at the University of Lille.

Valérie Viet Triem Tong was member of a recruitment committee for a Maitre de conférence position at the IUT de Vannes.

Valérie Viet Triem Tong was member of a recruitment committee for a Professor position at INSA Centre - Val de Loire.

Valérie Viet Triem Tong was member of a recruitment committee for a Professor position at INP ESISAR.

Gilles Guette was vice-president of a recruitment committee for a Maitre de conférence position at the University of Rennes 1.

9.2.1 Teaching

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,

• 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.