Chosen Plaintext Attack on Single Pixel Imaging Encryption via Neural Differential Cryptanalysis

Abstract Single pixel imaging (SPI) shows great potential in encryption by its indirect imaging mechanism. However, there appears to be room for further exploration in the corresponding cryptanalysis. Current studies primarily rely on straightforward end‐to‐end cryptanalysis of plain‐ciphertext pairs, ignoring the fundamental SPI optical path. As a result, the effectiveness of most attacks depends on the training data and the design of network, triggering low certainty and confidence. In this study, an alternative model is proposed to attack multiple SPI encrypting methods based on chosen plaintext attack framework, where arbitrary plaintexts can be encrypted as ciphertexts for cryptanalysis. In terms of the basic SPI setup, it is found that no matter how complicated the patterns are encrypted, the linear relationship between encrypted patterns and intensity always maintain. Thus, specifically, the ciphertext is first differentialized to derive encrypted patterns. By further reconstructing the pixel correlation of these derived patterns, deep learning is employed to correct them. Ultimately, the cracked patterns are used to decrypt plaintexts by conventional correlation. The experiments demonstrate that this method possesses a certain degree of reusability in the SPI encryption with linear propagating characteristic, like pattern‐encrypting class, demonstrating potential for the indirect optical encryption.