Cross subkey side channel analysis based on small samples.

Area of Science:

• Cryptography
• Machine Learning
• Computer Security

Background:

• Deep-Learning Side-Channel Analysis (DL-SCA) typically trains models on subkey-specific operations.
• Limited training data (e.g., 5K power traces) hinders deep learning model performance in DL-SCA.
• Existing methods struggle with insufficient data for effective feature learning.

Purpose of the Study:

• To propose and evaluate a novel cross-subkey training approach for Deep-Learning Side-Channel Analysis (DL-SCA).
• To enhance the effectiveness of DL models when dealing with limited power trace datasets.
• To reduce the data acquisition requirements for successful side-channel attacks.

Main Methods:

• Implemented a cross-subkey training strategy, augmenting data by training on SBox operations from multiple subkeys, not just the target.
• Utilized deep learning models for analyzing power traces from AES-128 implementations.
• Validated the approach on microcontroller (STM32F3) power traces and two public datasets.

Main Results:

• The cross-subkey training model achieved a significant accuracy improvement compared to individual subkey training.
• Demonstrated effectiveness on the STM32F3 microcontroller implementation of AES-128.
• Validation across multiple datasets confirmed the robustness and practicality of the proposed method.

Conclusions:

• Cross-subkey training is an effective trace augmentation technique for DL-SCA with limited datasets.
• The method significantly improves attack accuracy while reducing the need for extensive data collection.
• This approach offers a practical solution for real-world side-channel analysis scenarios.