PhD positions on Formal Methods for the Safety of Artificial Intelligence

Context

LIX (Laboratoire d'Informatique de l'Ecole Polytechnique) is a joint research unit with two supervisory institutions, École Polytechnique, a member of the Institut Polytechnique de Paris (cluster of universities composed of Ecole Polytechnique, Télécom Paris, ENSTA Paris, Télécom Sud Paris, ENSAE), and the Centre National de la Recherche Scientifique (CNRS), and one partner, Inria Saclay, with shared buildings and mixed teams.

LIX is organized in four poles: “Computer Mathematics”, “Data Analytics and Machine Learning”, “Efficient and Secure Communications”, “Modeling, Simulation and Learning” and “Proofs and Algorithms”. The Ph.D. student will be part of the Cosynus team in the “Proofs and Algorithms” pole. The members of the Cosynus team work on the semantics and static analysis of software systems, sequential, concurrent or distributed, and hybrid/control systems and cyber-physical systems.

The Ph.D. student will benefit from the exciting environment of LIX, specifically the Computer Science Department of Ecole Polytechnique (DIX), in which he or she will be able to give courses, and the Department of Computer Science, Data and Artificial Intelligence of Institut Polytechnique de Paris (IDIA). He or she will also interact with the project members of the SAIF project (Safe Artificial Intelligence through Formal Methods), of the French National Research Programme on Artificial Intelligence PEPR IA.

1st position: Probabilistic analysis of neural networks

AI is now embedded into a number of everyday life applications. More and more, we depend on neural networks to make even critical situations, such as control and motion planning for autonomous cars, and it is of primary importance to be able to verify their correct behavior. But for some applications, e.g. perception in autonomous systems, through classifiers, we can only hope for weaker properties such as probabilistic safety.
The context of this Ph.D. proposal is to provide qualitative measures of how likely a neural network classifier, or a neural network used in a control system, is to exhibit a certain behavior. A classical property for such classifiers is some form of robustness: for instance if an input image is slightly perturbed, the classifiers should classify this perturbed image the same way as the original one. The aim of this work is to tackle more general properties than local robustness, providing some explainability of the network behavior [RSBY+22].
The techniques that will be explored in the first part of the thesis will be based on imprecise probabilities, generalizing probabilistic and epistemic uncertainties (PBoxes and Dempster-Shafer structures [AAOB+13], polynomial forms [SSYC+20]), in the style of existing work such as [GFS+24] (for hybrid systems) and [EGSP22] (for neural-network based control systems). As a first step, this supposes defining good models to handle and propagate efficiently correlations in the Dempster-Shafer framework, possibly using copulas [Sch23].

[AAOB+13] Assalé Adjé, Olivier Bouissou, Jean Goubault-Larrecq, Eric Goubault, Sylvie Putot. “Static Analysis of Programs with Imprecise Probabilistic Inputs”. VSTTE 2013.
[EGSP22] Eric Goubault, Sylvie Putot, “RINO: Robust INner and Outer Approximated Reachability of Neural Networks Controlled Systems”. CAV 2022.
[GFS+24] Ander Gray, Marcelo Forets, Christian Schilling, Scott Ferson, and Luis Benet. “Verified propagation of imprecise probabilities in non-linear odes”. International Journal of Approximate Reasoning 2024.
[RSBY+22] Rabia Saleem, Bo Yuan, Fatih Kurugollu, Ashiq Anjum, Lu Liu. “Explaining deep neural networks: A survey on the global interpretation methods”. Neurocomputing, 2022.
[Sch23] Bernhard Schmelzer. “Random sets, copulas and related sets of probability measures”. International Journal of Approximate Reasoning. 2023.
[SSYC+20] Sriram Sankaranarayanan, Yi Chou, Eric Goubault, Sylvie Putot. “Reasoning about uncertainties in Discrete-Time Dynamical Systems using Polynomial Forms”. Neurips 2020.

2nd position: Set-based verification of quantified properties of neural networks

Abstraction-based safety verification for neural networks has received considerable attention recently, with in particular range analysis of neural networks using polyhedric abstractions (e.g. DeepPoly and deepZ [SGMPV2018,SGPV2019]) and reachability analysis of closed-loop systems with neural network controllers.
The objective of the thesis is to design sound abstractions to prove or falsify generalized robustness properties of neural network enabled systems. This includes in particular temporal properties of neural network-controlled cyber-physical systems and hyperproperties of neural networks, with applications in e.g. fairness analysis, or to compare the quality of neural networks obtained by sparsification and quantization to the baseline.
The tractable inner and outer-approximations of ranges of functions and neural networks proposed in [GP2020,GP2022] are a building block for proving very general quantified reachability problems [EGSP2024]. These constitute a basis from which the PhD student is expected to design new set-based methods to tackle these difficult properties of neural network controlled systems. Another possible starting point can be to explore some classes of polyhedras such as tropical abstractions [GPPRS2021] which are well-suited to abstract ReLU-activated neural networks.

[GP2020] Goubault, E., Putot, S.: Robust under-approximations and application to reachability of non-linear control systems with disturbances. IEEE Control Systems Letters 4 (4), 928–933 (2020)
[GP2022] Eric Goubault, Sylvie Putot: RINO: Robust INner and Outer Approximated Reachability of Neural Networks Controlled Systems. CAV (1) 2022: 511-523
[GPPRS2021] Eric Goubault, Sébastien Palumby, Sylvie Putot, Louis Rustenholz, Sriram Sankaranarayanan: Static Analysis of ReLU Neural Networks with Tropical Polyhedra. SAS 2021: 166-190
[SGMPV2018] Singh, G., Gehr, T., Mirman, M., Püschel, M., Vechev, M.T.: Fast and effective robustness certification. In: Advances in Neural Information Processing Systems, NeurIPS. pp. 10825–10836 (2018)
[TYMMN+2020] Tran, H.D., Yang, X., Manzanas Lopez, D., Musau, P., Nguyen, L.V., Xiang, W., Bak, S., Johnson, T.T.: NNV: The neural network verification tool for deep neural networks and learning-enabled cyber-physical systems. In: Lahiri, S.K., Wang, C. (eds.) Computer Aided Verification. pp. 3–17 (2020)
[SGMPV2018] Gagandeep Singh, Timon Gehr, Matthew Mirman, Markus Püschel, and Martin Vechev: Fast and Effective Robustness Certification. NeurIPS 2018.
[SGPV2019] Gagandeep Singh, Timon Gehr, Markus Püschel, and Martin Vechev: An Abstract Domain for Certifying Neural Networks. POPL 2019.
[RSBY+22] Rabia Saleem, Bo Yuan, Fatih Kurugollu, Ashiq Anjum, Lu Liu. “Explaining deep neural networks: A survey on the global interpretation methods”. Neurocomputing, 2022.
[GP2024] E. Goubault, S. Putot. “Inner and outer approximate quantifier elimination for general reachability problems“. HSCC 2024.

Competences

Candidates should have a strong background in computer science with an interest in formal methods and specifically abstract interpretation/set-based methods, artificial intelligence, and possibly the application to cyber-physical and controlled systems.

The working languages are English and French.

Conditions

Monthly gross salary of between 2100 and 2300 euros per month plus 232 euros extra in case of teaching (64 hours per year).

Extra perks include full social security coverage, partial reimbursement of public transport costs and subsidized meals.