Pseudo-Labeling for Deep Learning-Based Side-Channel Disassembly Using Contextual Layer and Feature Engineering

Abstract

Embedded devices face critical cyber-attacks due to their lightweight design and the sensitive data they handle. Integrating cloud and embedded systems increases the need for security measures against threats. Among these threats are deep learning-based side-channel disassembly attacks, which can expose sensitive information or steal software intellectual properties. Conducting a security test to evaluate the systems against these threats is essential. However, the main challenges include a comprehensive and refined dataset for training deep learning-based side-channel attacks and the lack of public datasets; labeling and profiling such attacks are costly and time-consuming. Additionally, accurately disassembling a single instruction is difficult due to the multiple classes representing each instruction and the obfuscation caused by dummy instructions. This study aimed to create an advanced side-channel evaluation methodology that performs three main deep-learning tasks: profiling using context-aware pseudo-labeling techniques at an instruction level, a disassembly model enhanced with moving log-transformed temporal interaction features, and a sequence labeling model for the detection of dummy instructions using natural language processing techniques. Utilizing gated recurrent units, the proposed pseudo-labeling model achieved 0.996 R2 in estimating the power trace for the assembly instructions. The proposed features improved the disassembly model's accuracy to 0.993, outperforming the related works. Additionally, the detection of dummy instructions using a long short-term memory model reached an accuracy of 0.979. This study provides valuable insights and methodology for measuring the software robustness against side-channel attacks.