Deep learning-based encryption scheme for medical images using DCGAN and virtual planet domain

Abstract

AbstractThe motivation for this article stems from the fact that medical image security is crucial for maintaining patient confidentiality and protecting against unauthorized access or manipulation. This paper presents a novel encryption technique that integrates the Deep Convolutional Generative Adversarial Networks (DCGAN) and Virtual Planet Domain (VPD) approach to enhance the protection of medical images. The method uses a Deep Learning (DL) framework to generate a decoy image, which forms the basis for generating encryption keys using a timestamp, nonce, and 1-D Exponential Chebyshev map (1-DEC). Experimental results validate the efficacy of the approach in safeguarding medical images from various security threats, including unauthorized access, tampering, and adversarial attacks. The randomness of the keys and encrypted images are demonstrated through the National Institute of Standards and Technology (NIST) SP 800-22 Statistical test suite provided in Tables 4 and 14, respectively. The robustness against key sensitivity, noise, cropping attacks, and adversarial attacks are shown in Figs. 15–18, 22–23, and 24. The data presented in Tables 5, 6, and 7 shows the proposed algorithm is robust and efficient in terms of time and key space complexity. Security analysis results are shown (such as histogram plots in Figs. 11–14 and correlation plots in Figs. 19–21). Information Entropy ($$7.9993 \pm 0.0001$$), correlation coefficient ($$\pm 0.09$$), Mean Square Error (MSE) ($$4166.3107 \pm 1645.2980$$), Peak Signal to Noise Ratio (PSNR) ($$12.2643 \pm 1.7032$$), Number of Pixel Change Rate (NPCR) ($$99.60\% \pm 0.2\%$$), and Unified Average Changing Intensity (UACI) ($$33.47\% \pm 0.1\%$$) underscore the high security and reliability of the encrypted images, are shown in Tables 8–11. Further, statistical NPCR and UACI are calculated in Tables 12 and 13, respectively. The proposed algorithm is also compared with existing algorithms, and compared values are provided in Table 15. The data presented in Tables 3–15 suggest that the proposed algorithm can opt for practical use.