Unlearning to Protect: A Distilled Reinforcement Learning Framework with Privacy-Preserving Feature Unlearning and XAI for IoT Security

Unlearning enables privacy protection without full retraining: The framework can selectively "forget" sensitive or outdated features from AI models through weight modification, allowing organizations to remove problematic data patterns while maintaining model performance—a capability absent in traditional AI systems that require complete retraining to address privacy concerns.

Lightweight AI models can be deployed on resource-constrained IoT devices: Using knowledge distillation, the researchers compressed a complex teacher model into a lightweight student model that achieved 99.60% accuracy in detecting botnets, making advanced AI-based security practical for IoT devices that typically lack computational power for heavy security solutions.

Explainability is critical for understanding security decisions: The integration of LIME (a technique for explaining AI decisions) helps security teams understand which features the model uses to identify threats, improving trust in automated security systems and enabling human oversight of AI-driven botnet detection.

Reversible unlearning demonstrates responsible feature removal: The framework showed that "restored features" maintained negligible performance loss after unlearning, suggesting that privacy-protective modifications can be carefully controlled and potentially reversed—important for maintaining model reliability while protecting sensitive information.

Comprehensive training data improves model robustness: By using 25% of the BoT-IoT dataset (compared to 5% in benchmark models), the research produced stronger detection capabilities, highlighting that responsible AI development requires adequate data representation rather than shortcuts that may compromise security effectiveness.