Bridging Lab and Industry: Practical SPA-GPT on Cryptosystems Boosted by LSTM and Simulated Annealing

Simple Power Analysis is a commonly used method in Side-Channel Analysis on cryptosystems, which requires a significant amount of labor costs for segmentation. General Pulse Tailor for Simple Power Analysis (SPA-GPT) proposed in CHES 2024 utilizes reinforcement learning to achieve automated segmentation. However, its low efficiency and only targeting public-key algorithms limit its practical applications. In this paper, we propose a practical method, which utilize long short-term memory network and attention mechanism, coupled with a new deep Q-network policy using Simulated Annealing strategy, to solve the contradiction between reinforcement learning and high efficiency in trace segmentation. Moreover, the novel agent proposed in this paper also demonstrates transferability, enabling direct segmentation of a trace under varying lengths and signal-to-noise ratio conditions once the agent has been fully trained. In addition, our new approach is applicable for locating each execution of block ciphers in various encryption modes. Comparative experiments are conducted on 14 datasets, which are collected from software or hardware implementations of RSA, ECC, ML-KEM, AES, PRESENT, and SIMON, running on microcontrollers, FPGAs, or smart cards. Experimental results show that the new method enhances time efficiency by 50.34% to 94.24% while reducing network parameters by 87.84% compared to SPA-GPT.