Reinforcement Learning-Enhanced Anti-Phishing Method with Adversarial Edge Collaboration for Smart Grid IoT Networks

To address the challenges of limited computing resources and high energy consumption in phishing detection for lightweight Internet of Things (IoT) devices, we propose WDPhishingDS, a distributed detection system enhanced by Reinforcement Learning (RL). This method adopts a hierarchical architecture: a pruned and RL-optimized lightweight model is deployed on IoT devices, achieving 90.3% accuracy for preliminary screening (18.7% higher than traditional static approaches) through dynamic threshold adjustments driven by [Formula: see text]-learning with parameters [Formula: see text], [Formula: see text], and [Formula: see text]. Suspicious links are securely transmitted to a smart grid IoT device via 5G/Bluetooth with encrypted RL-based adaptive routing, where a high-precision model performs secondary verification (97.4% accuracy, reducing false positives by 42% compared to single-stage methods). A Multi-Agent Reinforcement Learning (MARL) mechanism governs energy-aware task migration, enabling cross-device collaboration by evaluating real-time metrics like battery levels, processing loads, and latency. Experimental results show that RL-driven migration reduces energy consumption. Compared with static policies, energy consumption is reduced by 35.2%. Meanwhile, this migration maintains detection robustness against adversarial phishing tactics. This approach establishes an edge-coordinated security architecture, effectively countering AI-driven human-like phishing attacks and safeguarding smart grid IoT device user privacy through iterative RL optimization of resource allocation.