ECOLE CENTRALE DE LYON - Doctorant Analyse et synthèse de Détecteurs d’Anomalies sous contrainte de temps-Critique

Working environment

The École Centrale de Lyon (ECL) is a public scientific, cultural and professional institution. Mmember of the Ecoles Centrales group and the Écoles Nationales d’Ingénieurs network, ECL trains high-level generalist engineers, specialized engineers, masters students and doctoral candidates. The school hosts 2,500 engineering students and trainees, 300 master students and more than 250 doctoral students. It is characterized by recognized research supported by 6 research laboratories. ECL’s research activities are directed to and for the business world through numerous industrial contracts. 

The Ampère-lab is a joint research unit (CNRS, Ecole Centrale de Lyon, INSA Lyon, Université Lyon 1) of more than 150 researchers based in Lyon, France, working on the rational use of energy in systems in relation to their environment. The research carried out by the Automatique pour l’Ingénierie des Systèmes (AIS) department includes the development of methods

and tools for optimizing and controlling the dynamic behavior of systems in a wide range of

application domains, in collaboration with other departments of the laboratory and other engineering laboratories. The combination of theoretical and applied dimensions of this research

constitutes its great originality.

Over the last few years, the advisors have been working on the possibilities offered by Systems, Control and Signal approaches for the development of methods for the design/understanding of systems from different disciplines (electronics, electrical engineering, mechanics, biology, etc.).

In particular, expertise has been developed in the design of systems obtained by interconnecting subsystems, for which the combination of the input-output approach with (convex [BTN01, BV04]) optimization tools seems to be particularly effective. Convincing results have already

been obtained, ranging from upstream methodological contributions (e.g. [PKZS23,ACPKS23, LKD+17]) to their application to problems of strong practical interest (e.g. [PKS+21,KSCB16, GFS11]), and even to patent deposit (e.g. [PKZ+17, CGK10, CK13]). Recently, some team members have begun to explore the topic of the security of cyber-physical systems (e.g.,[PCZ21b,PCZ21a,EMSZ20]).

Scientific background of the project

The challenge of system security is to ensure that the specifications are met even in the face of malicious behavior or unforeseen events. Historically divided into the fight against physical attacks and the protection of information technologies, the significant increase in cyber-attacks against controlled systems (industrial infrastructures, power grids, drones, ... ) over the last two decades [DPF+19, SRE+19] and the limitations of conventional approaches have made it necessary to develop a systemic approach to system security [SAJ15], taking into account in particular the interaction between the cyber and physical worlds. On the one hand, traditional IT security methods focus primarily on protecting information

and do not directly consider the possible physical consequences of cyber-attacks. On the other hand, classical Control and Signal Processing approaches address tolerance to independent disturbances, but do not consider possible attacks by malicious rational actors. Thus, over
the last decade, approaches have been developed to prevent, detect, and mitigate attacks on controlled systems [CST19, DPF+19].

An important issue is to find an appropriate trade-off between the desired level of security and the satisfaction of a functional specification. This is due in particular to the difficulty of assessing the risk, and especially the likelihood, of an attack due to the heterogeneity of attackers, both in terms of objectives and resources [TSSJ15].

Problem and Objectives of the thesis
In this context, the ambition of this thesis is to tackle the problem of efficient analysis and synthesis of anomaly detectors (attacks, faults) under critical-time constraint. The critical-time is the maximum time horizon for which a system is considered to be safe after the occurrence of an anomaly, i.e. the system is not in a critical state and is still able to return to a normal mode (Fig. 4). This recently introduced security metric [PCZ21a]
seems to be relevant at every stage of the risk management process (analysis, prevention, detection, mitigation). The underlying motivation is that an increase in critical-time gives defense mechanisms, including human operators, more time to detect and mitigate anomalies.

The performance of anomaly detectors is traditionally evaluated according to three criteria : detection rate, false alarm rate and detection delay. When synthesizing a detector, only the first two criteria are considered. The estimation of the detection delay and the verification that the system does not enter a critical state before detection (i.e. that the detection delay is less than the critical time) are then estimated in a post-synthesis phase
using simulations. The first objective of this thesis is to develop an algorithmically efficient analysis method that allows to formally guarantee whether the worst-case detection delay of a given filter is less than the critical time. To achieve this, the research could build on previous work [PCZ21b,EMSZ20,BTMS17] on critical time computation and robust simulation techniques, and extend them to systems of practical interest (such as uncertain linear timevarying systems). The second objective is to propose a method able to take into account the critical-time constraint directly in the synthesis of the detection filter. An important challenge will be to obtain of an algorithmically efficient method, which may require to simplify
the problem (constraints relaxation, reformulation or relevant approximation of the problem, etc.). These first two objectives take place in the context of sharp anomalies, i.e. with a severe impact on the system in a reduced amount of time. The third objective will be to synthesize an anomaly detector in order to guarantee a minimum critical-time for the so-called stealthy attacks, i.e. attacks designed not to be detected by the filter.
As the expected contributions are mainly methodological, the results will be valorized mainly through presentations at international conferences and publications in leading journals in the field of Control and System theory

Candidate profile

Any candidate holding an engineering degree or a Master’s degree, with an excellent academic record, specialized in System and Control or Signal Processing OR a general degree with good skills in Applied Mathematics. An interest in the development of optimization-based methods and experience with Matlab would also be appreciated.

Recruitment process
Interested candidates, or those wishing more information, are warmly invited to send an e-mail containing a CV + a short message of presentation and motivation to the advisors team (see e-mail addresses at the beginning of this document).
The recruitment process consists of three stages :
1. Application until 05/05/2024. Oral interview by the advisors team and selection
of the candidate.
2. Oral interview by the EEA Doctoral School Board in late May/early June.
3. Final result : first half of June.