STAR: Self-Training Assisted Refinement for Side-Channel Analysis on Cryptosystems

Reconstructing cryptographic operation sequences through side-channel analysis is essential for recovering private keys, but practical attacks are hindered by unlabeled, noisy, and high-dimensional power traces that challenge accurate classification. To address this, we propose STAR, a two-stage unsupervised clustering correction framework. First, a Gaussian Mixture Model (GMM) performs an initial clustering to generate reliable pseudo-labels from high-confidence samples. Next, a self-training mechanism uses these pseudo-labels to train a Convolutional Neural Network (CNN), which then iteratively reclassifies low-confidence samples to refine the entire dataset. Validated on standard ECC, RSA, and SM2 datasets, our framework achieved 100% classification accuracy, demonstrating a significant improvement of 12% to 48% over state-of-the-art methods. These findings confirm that STAR is an effective and robust framework for enhancing the precision of unsupervised side-channel analysis, thereby strengthening key recovery attacks.