AutoProfile: Automated profiling in deep learning-based side-channel analysis.

Area of Science:

• Cryptography and Information Security
• Machine Learning Applications

Background:

• Side-channel analysis (SCA) exploits leaked information for data extraction from cryptographic systems.
• Deep learning (DL) shows promise for SCA, but network construction remains a challenge.

Purpose of the Study:

• To introduce AutoProfile, a novel methodology to improve DL-based profiling attacks on cryptographic systems.
• To enhance the efficacy of SCA by optimizing DL network selection for profiling attacks.

Main Methods:

• AutoProfile customizes the modeling strategy and acquisition function within Bayesian optimization for SCA.
• The methodology was evaluated using publicly available datasets with real side-channel measurements.
• Performance was compared against state-of-the-art methods on robust cryptographic targets.

Main Results:

• AutoProfile achieved an average performance enhancement of 78.4% over existing state-of-the-art methods.
• For targets with masking, random delay, and key variation countermeasures, AutoProfile drastically reduced required traces from thousands to dozens.
• AutoProfile demonstrated faster identification of effective DL networks compared to baseline methods across all tested SCA datasets.

Conclusions:

• AutoProfile significantly improves the efficiency and effectiveness of DL-based SCA.
• The method offers a substantial advantage in breaking robust cryptographic systems with minimal data requirements.
• AutoProfile provides a faster approach to selecting optimal DL networks for side-channel profiling attacks.