Adversarial Co-Evolution: Bilevel Optimization Defense Against RL-Adaptive Malware
arXiv·medium signal
Proposes a bilevel optimization framework where attacker and defender models co-evolve, replacing one-shot adversarial training that fails against adaptive RL-based malware generators. The defense continuously adapts to increasingly sophisticated evasion strategies rather than training against a static threat model. Addresses the arms-race dynamic where traditional defenses degrade as attackers learn to bypass them.