Cryptanalyzing an Image Encryption Algorithm Underpinned by a 3-D Boolean Convolution Neural Network
You Ke, Pengbo Chen, Chengqing Li
Abstract
This article analyzes the security performance of an image encryption algorithm based on a 3-D Boolean convolutional neural network (CNN). The algorithm utilizes the convolutional layers of a CNN as the encryption component, thereby achieving low-precision computations. The convolution matrices and kernels required in the convolutional layer are generated using a prime modulo multiplicative linear congruence generator. The permutation part is used to confuse the cipher image obtained from the convolution of the encryption algorithm. However, due to its low-precision computation, this encryption algorithm employs one-to-one XOR and modulo operations, altering individual pixel values exclusively during encryption without diffusing changes to neighboring pixels. Capitalizing on this vulnerability, we propose chosen-plaintext attacks on the one-round and multiple-round versions of this encryption algorithm. Using a divide-and-conquer strategy, the convolution and permutation parts of the encryption algorithm are attacked separately.