Optimization of Biomedical Image Encryption Algorithm Based on Chaos-CNN Hybrid Network

Liping Liu
School of Cyberspace Security, Gansu University of Political Science and Law, Lanzhou 730070, Gansu, China.

Abstract:

In response to the security issues of sensitive data transmission in biomedical imaging, this essay proposes a medical image encryption framework (CCE) based on a chaotic-convolutional neural network (CNN) hybrid structure, aimed at addressing the energy consumption and security issues of traditional cascade encryption methods in the secure transmission of medical data. The SHA3-512 hash function is used to extend the user key to generate a dual-channel dynamic key, in which one activates the multi-mode chaotic system to inhibit digital degradation, and the other uses the CNN weight self-adjustment mechanism to construct a nonlinear mapping relationship, which improves the key space by 3 orders of magnitude compared with the traditional method. The Swin Transformer model is introduced to analyze the block-level entropy of the image, and an adaptive matching mechanism of entropy value-quantization parameter-encryption strength is established, which reduces the encryption energy consumption by 37% on the basis of maintaining more than 96% of the local structural integrity. The designed hybrid measurement matrix satisfies the limited equidistant nature of compressive sensing through Tikhonov regularization constraint number and Gram-Schmidt orthogonalization processing, and improves the signal reconstruction accuracy by 23% compared with the traditional Toeplitz matrix. By constructing a ciphertext mutual information obfuscator through the adversarial generation network, the success rate of known plaintext attacks is reduced to 0.4%, and the detection accuracy of selected ciphertext attacks exceeds 98%. Experiments show that the reconstructed PSNR of the framework on the GPU/FPGA/CPU hybrid platform reaches 42.6dB (JPEG 36.4dB), and the system energy efficiency is increased by 58% compared with AES+JPEG, and the contribution of each module is verified by ablation experiments. Through the dynamic randomness of chaotic systems and the multi-scale feature extraction capabilities of CNNs, this framework can adaptively distinguish sensitive areas in medical images and allocate differential encryption strengths to them, thereby enhancing security while ensuring the integrity of critical diagnostic information.