TPE-ADE: Thumbnail-Preserving Encryption Based on Adaptive Deviation Embedding for JPEG Images

The growing practice of outsourcing captured photos to the cloud has provided users with convenience while also raising privacy concerns. Traditional image encryption techniques prioritize privacy protection but often compromise usability, which is unacceptable for cloud users. To strike a balance between image privacy and usability, scholars have proposed thumbnail-preserving encryption (TPE), whose cipher image preserves the same thumbnail as the plain image while erasing details beyond the thumbnail, providing visual usability while protecting privacy. Regrettably, most of the proposed TPE schemes are not well-suited for widely used JPEG images, and existing TPE schemes supporting JPEG suffer from drawbacks such as poor visual usability, high expansion rate, and the inability to decrypt without loss. Besides, the retrieval designed for TPE-encrypted images exhibits limited generalization. To address these challenges, we pertinently introduce a TPE based on adaptive deviation embedding (TPE-ADE) for JPEG images, incorporating Huffman coding and reversible data hiding techniques. By leveraging JPEG in-compression encryption, we achieve perfectly reversible TPE that enhances visual usability and reduces expansion rates of TPE-encrypted images. Additionally, we encourage the TPE-encrypted images to resemble low-resolution images (LRIs). Then, the convolutional neural network (CNN) is employed to recognize and retrieve LRIs to verify the functionality of TPE-encrypted images. Also, a teacher-assistant-student (TAS) learning paradigm is proposed to optimize the CNN model, enhancing the performances of recognition and retrieval. Experimental results validate the superiority of our encryption algorithm and the effectiveness of TAS.