DeepEnc: deep learning-based CT image encryption approach

The rapid development of artificial intelligence gave motivations to researchers to apply it in information security, biometrics, and medical image encryption. This paper presents an encryption approach for CT images that are used in diagnosing COVID-19 disease. To add more privacy, the proposed encryption mechanism links the CT image and the facial image of a person. Firstly, a simple Convolutional Neural Network (CNN) is utilized to extract features from facial images. The architecture of the CNN model includes three convolutional groups, and each group consists of a convolutional layer, a batch normalization layer, a Rectified Linear Unit (ReLU), and a maximum pooling layer. Moreover, the extracted feature vector is utilized to adjust the initial states of a two-dimensional Sine Logistic Modulation Map (2D-SLMM), so as to generate chaotic matrices SLM1 and SLM2, and to adjust the initial states of a Tent Logistic Map (TLM) in order to get two other chaotic matrices, TLM1 and TLM2. Finally, the Chaotic Magic Transform (CMT) confusion operation and the bitwise XOR diffusion operation are used to randomly scramble the pixel locations and change the pixel values, respectively. Only after two rounds of CMT and two bitwise XOR operations with chaotic matrices SLM1, SLM2, TLM1 and TLM2, we get an unrecognizable encrypted CT image with a high security level. In adittion, we study and analyze the performance of the proposed approach in the presence of differential, entropy and key sensitivity attacks. In the experiments, when performing decryption, two cases are discussed; utilization of the correct key and utilization of the correct key with small variation. The robustness of the proposed encryption approach is measured using different metrics like Correlation Coefficient (CC), entropy, histogram analysis, and elapsed time. The maximum elapsed time is about 0.4361 seconds.