Explainable AI Based Diagnosis of Poisoning Attacks in Evolutionary Swarms

Swarm systems are vulnerable to data poisoning attacks: Multi-agent systems like drone networks that coordinate autonomously can be compromised through data manipulation, leading to breakdowns in cooperation or adversarial behavior among agents—a critical concern for real-world applications in disaster response and surveillance.

Poisoning effects become detectable at relatively low thresholds: The research found that when poisoned data reaches just 10% of the system's inputs, explainable AI methods can identify resulting inefficient cooperation and non-optimal strategies, suggesting vulnerabilities emerge sooner than might be expected.

Explainable AI enables diagnosis of attacks: Rather than treating poisoned swarms as black boxes, the researchers used explainable AI techniques to extract interpretable "footprint characterizations" that help identify when and how attacks have compromised team coordination—making it possible to diagnose problems rather than just detect failures.

Understanding attack mechanisms is essential for defense: By systematically studying how data poisoning affects swarm behavior through evolutionary intelligence models, the framework provides insights into attack patterns that could inform better safeguards for autonomous multi-agent systems operating in critical environments.