DeepKey-TPE: Using Deep Learning to Generate Key for Image Security With Continued Usability Through Thumbnail-Preserving Encryption

Recently, digital images are the most important information carrier of data obtained from consumer devices, thus playing a vital role in various potential applications. Despite benefits of these images, it still faces information leakage and other significant challenges when it comes to ensuring formal privacy guarantees, especially in consumer devices where user data is highly sensitive. The leakage of such sensitive data can have severe consequences. Traditional encryption methods focus on high security but often ignore usability, making them less practical when privacy and accessibility are needed simultaneously. This study proposes a secure model, called DeepKey-TPE, for image encryption based on deep learning-based key generation for high security using thumbnail-preserving encryption (TPE). The digital image is initially segmented into sensitive and non-sensitive regions using multi-task cascaded convolutional neural network (MTCNN). Sensitive areas are then encrypted using a deep learning-based encryption method. However, managing the image’s privacy and usability in cloud environments is more effective when using the transformation-based TPE method to encrypt the non-sensitive regions. Experimental results demonstrate that DeepKey-TPE achieves high key randomness with low autocorrelation, producing encrypted images with an entropy of 7.999, NPCR of 99.60%, and UACI of 33.40% in just 0.728s, thereby outperforming state-of-the-art schemes in terms of security, efficiency, and preservation of user privacy and image usability.