Efficient AES Side-Channel Attacks Based on Residual Mamba Enhanced CNN.

Area of Science:

• Cryptography and Security Engineering
• Artificial Intelligence and Machine Learning
• Computer Architecture and Embedded Systems

Background:

• Deep learning methods are increasingly vital for side-channel attacks (SCA) due to their advanced feature extraction. Traditional Convolutional Neural Networks (CNNs) show limitations in modeling long-range sequential data, impacting attack efficiency.
• The need for more effective deep learning models in SCA is driven by the continuous advancement of sophisticated attack techniques.

Purpose of the Study:

• To propose a novel hybrid deep neural network architecture for improved modeling of side-channel information in Advanced Encryption Standard (AES) implementations.
• To enhance the efficiency and generalization capabilities of deep learning-based side-channel attacks.

Main Methods:

• A hybrid deep neural network integrating Residual Mamba blocks and Multi-Layer Perceptrons (MLP) was developed.
• The Residual Mamba module utilizes state-space modeling for capturing long-range dependencies and improving global temporal perception.
• The MLP module was employed for further fusion of high-dimensional features.

Main Results:

• The proposed model achieved guessing entropy (GE) rank 1 with fewer than 100 attack traces on the ASCAD dataset for AES.
• Experimental results demonstrate superior performance compared to traditional CNNs and Transformer-based models.
• The model exhibited fast convergence and high attack efficiency.

Conclusions:

• The developed hybrid deep neural network offers an effective new paradigm for deep learning in side-channel analysis.
• The findings have significant theoretical and practical implications for securing cryptographic implementations against advanced side-channel attacks.
• The proposed architecture enhances the modeling of sequential information, leading to more efficient and generalized SCA.