Improving deep learning-based neural distinguisher with multiple ciphertext pairs for speck and Simon.

Area of Science:

• Cryptography
• Machine Learning
• Computer Science

Background:

• Neural network-based differential distinguishers offer efficient cryptanalysis but face accuracy limitations in reduced-round cryptosystems.
• Gohr's 2019 neural distinguisher established a baseline, yet improvements are needed for practical security analysis.

Purpose of the Study:

• To propose a novel neural distinguisher with enhanced accuracy and efficiency for cryptanalysis.
• To investigate the impact of network architecture and dataset optimization on distinguisher performance.

Main Methods:

• Designed a new neural distinguisher incorporating multi-scale convolutional blocks and dense residual connections.
• Developed a novel dataset model by combining ciphertext pairs, their differences, keys, and key differences, informed by linear attack concepts.
• Conducted ablation studies to validate the efficiency of the proposed components.

Main Results:

• The proposed neural distinguisher achieved 0.15-0.45% higher accuracy than Gohr's for Speck 32/64 with single ciphertext pairs.
• With multiple ciphertext pairs, accuracy improved by 1.24-3.5% for Speck 32/64 and 0.32-1.83% for Simon 32/64 compared to existing methods.
• A key recovery attack using the novel distinguisher achieved a 61.8% success rate, outperforming Gohr's by 9.7%.

Conclusions:

• The novel neural network architecture and dataset model significantly enhance the accuracy of differential distinguishers.
• The proposed method demonstrates superior performance in cryptanalysis of block ciphers like Speck and Simon.
• This advancement offers improved security analysis capabilities through more effective differential cryptanalysis.