Deep learning-based improved side-channel attacks using data denoising and feature fusion.

Area of Science:

• Computer Science
• Cryptography
• Machine Learning

Background:

• Deep learning excels in side-channel attacks but faces challenges like complexity and noise sensitivity.
• Existing models often increase computational load and limit feature extraction for improved accuracy.
• Noise in datasets reduces data correlation, hindering the effectiveness of deep learning-based attacks.

Purpose of the Study:

• To propose novel deep learning architectures for more efficient and accurate side-channel attacks.
• To develop a denoising model to mitigate the impact of noise on attack performance.
• To evaluate the effectiveness of the proposed models on standard datasets.

Main Methods:

• An InceptionNet-based network structure was applied for side-channel attacks, featuring fewer parameters and parallel processing for faster convergence and efficiency.
• A LU-Net-based network structure, incorporating an encoder-decoder design with LSTM layers and skip connections, was developed for denoising.
• Experimental evaluations were performed using the ASCAD and DPA Contest v4 datasets.

Main Results:

• The InceptionNet-based attack model achieved high efficiency, requiring only 30 traces for key recovery on the ASCAD dataset and 1 trace on the DPA Contest v4 dataset.
• The LU-Net denoising model effectively reduced noise, preserving signal characteristics and improving overall attack performance.
• The proposed deep learning methods significantly enhance side-channel attack capabilities compared to traditional approaches.

Conclusions:

• The proposed InceptionNet and LU-Net models offer a more efficient and robust solution for deep learning-based side-channel attacks.
• These models address the limitations of existing methods by reducing complexity and improving noise resilience.
• The findings demonstrate a significant advancement in the field of cryptographic side-channel analysis.