Side-channel attacks represent a significant threat to information security by exploiting physical leakage such as power consumption, electromagnetic emissions, and temporal information generated during cryptographic algorithm execution to compromise secret keys. To address challenges in side-channel analysis, particularly limited physical access to targets for prolonged data acquisition and insufficient sample sizes, deep learning has been effectively integrated into side-channel modeling attacks. These neural network-based approaches demonstrate superior performance compared to classical template attacks, achieving key recovery with fewer attack traces. While data augmentation techniques – methodologies designed to enhance training dataset size or quality – have been adapted to side-channel contexts to improve attack efficacy, existing implementations often neglect the strong temporal correlations between sampling points within power traces. This study proposes a BiLSTM-based TimeGAN framework for dataset augmentation that preserves critical temporal characteristics of power traces. The statistical consistency between synthetic and original data is rigorously validated through Kolmogorov-Smirnov test. Experimental results demonstrate a 23.3% reduction in required attack traces compared to baseline methods (from 3,000 to 2,300 traces), substantiating the method’s effectiveness in conducting successful cryptographic attacks under limited-sample dataset conditions.