Measure the Psychometric Functions of Deep Learning Models in Encrypted Image Recognition Tasks

The research aims at applying the convolutional neural network (CNN) including LeNet5, AlexNet, and Visual Geometry Group (VGG) with 16 weight layers to directly classify among 4 categories of fully encrypted images that were encrypted by various cryptographic algorithms involving Advanced Encryption Standard (AES), Blowfish, Data Encryption Standard (DES), and Triple DES (TDES) into without decryption to establish a secure image querying technique. The investigation was implemented with three concrete tasks. Firstly, used CNN models to recognize enciphered images with different encryption algorithms or cryptographic keys. Secondly, applied CNN models to classify enciphered images with simulated inference of Gaussian noise. Thirdly, employed CNN models to identify encrypted images with simulated inference of the reduction of contrast ratios. The results achieved secure image recognition and proved the underlying capability of the CNN models to recognize encrypted images even with the interference of Gaussian noise and lower contrast ratio, which mostly are impossible for human beings with normal vision to distinguish. Furthermore, the study initially showed that the increased cryptographic strengths of encrypted images usually caused an implicit impact on the accuracy of the three CNN models. Conversely, the variations in the severity of Gaussian noise on enciphered images and the contrast ratio of encrypted images could have explicit impacts on the accuracy of the models. The results also reflected that LeNet-5 is the most suitable CNN model for recognizing enciphered images with different encryption strengths and recognizing enciphered images with Gaussian noise. Moreover, all three CNN models could be suitable when analyzing encrypted images with reduced contrast ratio according to the degree of reduction of the contrast ratio.