Operating-Regime Evaluation of Byzantine-Resilient Multi-Agent Reinforcement Learning for Sensor-Networked Safe...

Area of Science:

• Cyber-Physical Systems
• Multi-Agent Reinforcement Learning (MARL)
• Networked Systems Security

Background:

• Corrupted sensor data in networked cyber-physical systems compromises formation accuracy and safety.
• Byzantine-resilient multi-agent reinforcement learning (MARL) is crucial for maintaining system integrity.

Purpose of the Study:

• To evaluate the operating regimes of a trust-based safety pipeline (RS-MARL) under various Byzantine attack scenarios.
• To identify supported, inconclusive, and detector-limited performance regimes, rather than proposing a new MARL algorithm.

Main Methods:

• A multiplicity-corrected operating-regime protocol applied to RS-MARL.
• A large-scale evidence base including 3000 core runs, 580 control/ablation runs, and 2380 review-audit extension runs.
• Analysis covered five methods, six attack families, five Byzantine ratios, and specific audits like A-CBF calibration and sensor impairment.

Main Results:

• RS-MARL showed fewer safety violations than Safe-MAPPO in many scenarios, but differences were not statistically significant after correction.
• Detection was reliable against collusive, random, and stealthy attacks.
• Detection was weak or undefined against constant, adaptive, and sign-flip attacks, indicating limitations of the energy-based trust detector.

Conclusions:

• The study defines specific operating regimes for RS-MARL, highlighting its strengths and weaknesses under different Byzantine attack types.
• A reproducible reporting template for MARL safety evaluations is proposed, emphasizing matched baselines, sensitivity, reliability, and trade-offs.