Byzantine-resilient cooperative control of multi-agent systems: a three-layer defense framework

Malicious or faulty agents in cooperative multi-agent control systems pose significant challenges to system-wide performance through dissemination of corrupted information. This dissertation presents an integrated three-layer defense architecture that addresses Byzantine fault tolerance through complementary approaches: theoretical guarantees, real-time mitigation, and intelligent identification. The first defensive layer implements ell_1 convex optimization combined with Hankel matrix reconstruction, establishing provable recovery bounds under sparsity constraints. The second layer introduces RCP-f (Resilient Consensus Protocol with f-filtering), a novel distance-based filtering mechanism achieving O(n \log n) computational complexity while maintaining real-time operation. The third layer employs LSTM neural networks augmented with Correntropy-based statistical features, achieving 99\% classification accuracy in identifying compromised agents. Experimental validation employs an 8-agent heterogeneous cart-pendulum system. Results demonstrate that Byzantine attacks cause 4,976-fold performance degradation without countermeasures. The proposed framework achieves complete performance restoration (tracking error: 0.049 versus baseline: 0.048) with sub-millisecond computational latency. Comparative analysis across six experimental scenarios validates the complementary nature of the three layers: ell_1 optimization provides theoretical validation but lacks real-time applicability; RCP-f enables autonomous performance recovery; LSTM adds diagnostic capability without compromising control performance. This work represents an original contribution in systematically integrating data-driven optimization, real-time filtering algorithms, and machine learning methodologies for Byzantine-resilient cooperative control. Applications include autonomous vehicle platoons, UAV formations, and distributed industrial control systems.