Post-quantum cryptography (PQC) offers resistance against quantum adversaries. This study's practical implementations remain vulnerable to side-channel attacks (SCAs) that exploit timing, power, or electromagnetic leakage. In this study, we introduce an unsupervised, resource-efficient anomaly detection framework tailored to the unique constraints of post-quantum cryptography (PQC) systems. Unlike traditional methods that rely on labeled attack traces or algorithm-specific profiling, our approach leverages an autoencoder trained solely on benign traces to learn deep latent representations of normal cryptographic behavior. The system flags deviations using reconstruction error and supports multiple PQC schemes, including Kyber and Dilithium, without retraining. Experimental results demonstrate an average classification accuracy of 98.1%, with a false positive rate of 0.7% and a false negative rate of 0.4%. Under adversarial perturbation and Gaussian noise, the model maintains an AUC-ROC of 1.00, confirming its robustness. Additionally, ablation studies across CNN, GRU, and Transformer architectures validate the autoencoder’s superior trade-off between accuracy and latency, achieving an inference time of 0.036 ms and a model size of only 0.11 MB. This enables real-time deployment on constrained devices without sacrificing security. The proposed solution marks a step forward in scalable, adaptive post-quantum defenses and opens new directions for cryptographic anomaly detection with minimal overhead. This framework is deployable on real-world PQC-enabled IoT systems.
Praditasari et al. (Thu,) studied this question.