Improving Differential-Neural Cryptanalysis

Improving Differential-Neural Cryptanalysis

2024-09-02 | Liu Zhang, Zilong Wang, Baocang Wang
The paper aims to enhance the capabilities of differential-neural cryptanalysis by improving the accuracy and handling more rounds of encryption. The authors introduce an Inception module, which uses multiple parallel convolutional layers with varying kernel sizes before the residual block, to capture multi-dimensional information. They also expand the convolutional kernels in the residual blocks to increase the network's receptive field. These improvements lead to significant accuracy enhancements for SPECK32/64 and SIMON32/64, enabling the training of 9-round and 12-round differential-neural distinguishers, respectively. To ensure successful key recovery attacks, the authors use neutral bits to generate multiple plaintext pairs with the same difference, which are then encrypted to produce multiple ciphertext pairs. They also redefine the time complexity formula based on the prediction speeds of the distinguisher on a single-core CPU and a GPU. This allows for a more equitable comparison of results. The paper demonstrates the effectiveness of these improvements through key recovery attacks on 13-round SPECK32/64 and 16-round SIMON32/64, reducing both time and data complexity. Additionally, they achieve a 14-round key recovery attack on SPECK32/64 by exhaustively guessing a 1-round subkey, and a 17-round attack on SIMON32/64. Knowledge distillation techniques are used to reduce the model size, further accelerating the prediction speed and reducing time complexity.The paper aims to enhance the capabilities of differential-neural cryptanalysis by improving the accuracy and handling more rounds of encryption. The authors introduce an Inception module, which uses multiple parallel convolutional layers with varying kernel sizes before the residual block, to capture multi-dimensional information. They also expand the convolutional kernels in the residual blocks to increase the network's receptive field. These improvements lead to significant accuracy enhancements for SPECK32/64 and SIMON32/64, enabling the training of 9-round and 12-round differential-neural distinguishers, respectively. To ensure successful key recovery attacks, the authors use neutral bits to generate multiple plaintext pairs with the same difference, which are then encrypted to produce multiple ciphertext pairs. They also redefine the time complexity formula based on the prediction speeds of the distinguisher on a single-core CPU and a GPU. This allows for a more equitable comparison of results. The paper demonstrates the effectiveness of these improvements through key recovery attacks on 13-round SPECK32/64 and 16-round SIMON32/64, reducing both time and data complexity. Additionally, they achieve a 14-round key recovery attack on SPECK32/64 by exhaustively guessing a 1-round subkey, and a 17-round attack on SIMON32/64. Knowledge distillation techniques are used to reduce the model size, further accelerating the prediction speed and reducing time complexity.
Reach us at info@study.space