CDA-Cache: A Class-Aware Knowledge Distillation Approach with Collaborative Caching for Intrusion Detection on IoT Edge Gateways

Deploying intrusion detection models on resource-constrained IoT edge gateways is challenged by model complexity, class-imbalanced traffic, and redundant inference across gateways. CDA-Cache integrates a compact convolutional student, EQLv2-inspired gradient-balanced hard-label supervision with conventional temperature-scaled knowledge distillation, and quantized inference-result caching. Class-aware weights act on the hard-label branch rather than the KL-distillation term. The reported within-dataset results reach 99.43% Accuracy and 99.16% Macro-F1 on Edge-IIoTset and 93.67% Accuracy and 88.75% Macro-F1 on TON_IoT. The student uses 43,087 parameters, approximately 87.9% fewer than the CANET teacher, and has lower local inference cost on the evaluated Intel CPU. Finite logically shared-cache simulations indicate opportunities to reuse predictions, with hit rates of 53.95–94.19% across 2–20 simulated gateways. These results characterize the evaluated training recipe and reuse scenarios. CDA has not been isolated from SVMSMOTE, and comparative per-class results for the four training variants are unavailable. The reported cache data do not establish collision safety or configuration robustness; physical-gateway RAM, power, and energy were not measured.