Acoustic and vibration side channel analysis on post-quantum cryptography using image-based deep learning

Abstract

Post-quantum cryptography (PQC) is designed to resist quantum-era attacks; however, practical implementations remain vulnerable to physical side channel leakage. This work proposes an image-based acoustic–vibration side-channel analysis framework to assess non-invasive leakage in PQC systems. Acoustic and vibration signals from secret-dependent executions are modeled and transformed into time–frequency spectrograms using short time fourier transform (STFT). The dataset comprises 1,545 samples (1,236 training and 309 testing), acquired at 16 kHz and segmented into 2.5-second windows. Leakage classification is performed using convolutional neural networks (CNNs) and vision transformers (ViTs) under single-modality and multimodal fusion settings. Results show that acoustic signals yield strong leakage, achieving up to 100% accuracy with CNN, while vibration signals reach up to 98.75%. Multimodal fusion improves training stability and overall performance, and ViT models demonstrate better generalization across modalities. These findings confirm that multimodal spectrogram based deep learning is effective for PQC side-channel analysis and underscore the need for rigorous physical security evaluation in real-world PQC implementations.