Stability-Constrained Meta-Reinforcement Defence against Learning-Enabled Autonomous Malware

The rapid evolution of learning-enabled autonomous malware renders traditional intrusion detection systems ineffective by violating stationary threat assumptions. This paper introduces the stability-constrained meta-reinforcement defence, a framework modelling attacker–defender interaction as a partially observable stochastic game. Our methodology employs three key techniques: a Lyapunov-inspired stability constraint with formal proof of bounded escalation, online adversarial drift estimation via Kullback–Leibler (KL) divergence for dynamic threshold calibration, and regret-aware meta-learning with sub-linear regret guarantees. We implemented the framework in PyTorch and evaluated it on five benchmark datasets comprising over 19 million instances. Comparative analysis against 12 state-of-the-art baselines demonstrates our framework’s superiority, achieving accuracy on one dataset and on another with a false positive rate. Under adaptive attacks, our framework limits performance degradation to versus for the best baseline. Cross-dataset evaluation shows models trained on the most comprehensive dataset achieving average accuracy. Statistical significance tests confirm all improvements with p-values less than 0.001, validating that integrating Lyapunov stability constraints, drift estimation, and regret–aware meta-learning enables provably robust autonomous cyber defence against adaptive malware.