PhD Oral Exam - Seyedeh Mahshid Rahimifard, Electrical and Computer Engineering

When studying for a doctoral degree (PhD), candidates submit a thesis that provides a critical review of the current state of knowledge of the thesis subject as well as the student’s own contributions to the subject. The distinguishing criterion of doctoral graduate research is a significant and original contribution to knowledge.

Once accepted, the candidate presents the thesis orally. This oral exam is open to the public.

Abstract

Multi-agent systems (MASs) coordinate autonomous agents over communication networks. In the leader-following approach, a leader generates a reference trajectory that followers track by exchanging state information with their neighbours. The sensor-to-controller, controller-to-actuator, and inter-agent communication channels are attack surfaces. 

This thesis develops cyberattack detection and resilient control methods for discrete-time MASs with unknown-but-bounded noise, using set-membership methods that guarantee the true state lies inside a computed set at every time step.

The first contribution detects cyberattacks in nonlinear leader-following MASs using two-step fuzzy set-membership filtering. A Takagi-Sugeno fuzzy model approximates the agent nonlinearities over the true state, eliminating the base-point error. Two criteria detect and isolate attacks on the sensor, control, and communication channels, each followed by a matching recovery mechanism.

The second contribution unifies zonotopic estimation, consensus control, detection, isolation, and recovery for linear MASs. A free-weighting matrix, obtained from a convex offline program, reduces the conservatism of the outer-approximation zonotope. Three interface-specific tests detect and isolate attacks without false alarms, and the isolation outcome selects the matching recovery
action.

The third contribution addresses nonlinear MASs under simultaneous denial-of-service attacks on the control channels and deception attacks on the sensors. A two-step resilient set-membership estimator guarantees that every follower’s state lies in its prediction, updated estimation, and leader-state ellipsoids.

The fourth contribution removes the outer approximation of the preceding estimators. For discrete-time linear systems, the exact set of consistent states is represented as a standard zonotope in an augmented space. A generator-only pivot rule allows the pivot sets to be precomputed offline, and the estimate size is bounded in proportion to the plant convergence rate. An alarm is raised when
the exact set is empty, with zero false alarms.

The fifth contribution targets replay attacks, which no passive detector can identify in steady-state. A private multi-tone watermark is injected in each agent’s control input, and a necessary and sufficient detectability condition is established. A watermark with a period longer than the recording horizon and a sufficiently large amplitude ensures that replays at every offset within that horizon are detected, while the online controller gain preserves leader-following consensus.