OCRN-TransUNet: Enhancing one-to-many temporal optical data cryptanalysis with robustness against multiplicative noise

Abstract

To ensure the vigorous development of the data trading market, data circulation through optical transmission systems assumes paramount importance due to their exceptional transmission efficiency. Securing these systems is an imperative factor, and optical data cryptography offers distinct advantages in ensuring the security of data transmission within them. Furthermore, optical cryptanalysis serves as a catalyst for advancing optical cryptography technology, wherein the correctness of cracking, particularly in the presence of multiplicative noises, holds significant value. Differing in traditional optical cryptanalysis techniques, which incur substantial labor costs, deep learning-based optical cryptanalysis techniques offer a promising alternative that significantly reduces manual effort. However, current deep learning-driven optical cryptanalysis approaches are unable to crack one-to-many temporal optical data encryption methods due to the incomplete considerations within their feature extraction and loss function design. Additionally, they have rarely addressed the enhancement of model robustness against multiplicative noises, given the overt adverse influence that multiplicative noise exerts on cryptanalysis effectiveness. To overcome this problem, we introduce an optical cryptanalysis paradigm named OCRN-TransUNet, which harnesses the strengths of convolutions and transformers, thereby exhibiting immense potential in this challenging domain. Specifically, we combine sub-pixel convolutional layers to enhance TransUNet's capability in performing data upsampling tasks. Moreover, we devise a multifaceted loss function that includes binary cross entropy loss, Pearson correlation coefficient (PCC) loss, and an emendation loss term to bolster the reconstruction ability and robustness of the model. Utilizing plaintext and ciphertext pairs generated by the one-to-many temporal optical data encryption method, we train our deep learning model, achieving direct conversion from the ciphertext to plaintext without knowing the secret key. Experimental results show that the proposed method not only effectively cracks the encryption of one to multiple time domain optical data, but also outperforms TransUNet in terms of convergence as well as cryptanalysis ability and possesses certain robustness against multiplicative noise pollution. Specifically, in the absence of multiplicative noises, we observe an improvement in the Peak Signal-to-Noise Ratio (PSNR) by 1.0881, and an increase in the PCC by 0.0468. In the presence of multiplicative noises with a variance of 1, these enhancements are even more significant, with a boost in PSNR by 1.8337 and an increase in PCC by 0.1605.