Answers are grounded in this paper and cite their sources.
Computational study demonstrates robust multi-class cache side-channel attack detection using a hybrid neural network, highlighting improved defense against hardware-level data theft.
Cache Side Channel Attacks (CSCAs) leverage the resource-sharing feature inherent in modern multicore architectures to steal any confidential information, such as encryption keys, from victim programs. Current state-of-the-art detection approaches based on hardware performance counters (HPCs) depend greatly on static configurations of counters, work predominantly in binary mode, have no robustness evaluation to adversarial perturbations, and experience significant data leakage at the temporal level. In this paper, we propose a holistic leak-free multi-class approach to detect and classify CSCAs in four major categories: Flush + Reload, Prime + Probe, Flush + Flush, Evict + Time, along with benign execution. The first one is that the hybrid CNN-LSTM network with 240,837 parameters and 44 HPC features can reach 78.04% test accuracy with zero leaks in multi-class classification and 84.84% validation accuracy under the strict temporal block splitting scheme. It scores 77.29% 0.97% mean test accuracy on 30 runs. With realistic production priors, the model can score 96.0% accuracy rate, 100.0% attack detection sensitivity, and 0.8574% false positive rate. The second one is the evaluation of adversarial robustness with FGSM and 10-step PGD, where the model can score 62.81% accuracy even with perturbation. Finally, the evaluation of MI-based feature selection with redundancy removal is done through a six-variant ablation study.
No takes yet. Share an insight, caveat, or question.
Moon et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: