Encrypted JPEG Image Retrieval via Huffman-code Based Self-Attention Networks

Image retrieval has been widely used in daily life. In recent years, with the increasing awareness of privacy protection, encrypted image retrieval has also been gradually developed. In this paper, we propose a new encrypted JPEG image retrieval scheme, named Huffman-code Based Self-Attention Networks (HBSAN), which could conduct image retrieval and protect image privacy effectively. To be specific, we first extract Huffman-code histograms directly from cipher-images which are encrypted by jointly using new orthogonal transformation, permutation cipher and stream cipher during JPEG compression. Then we employ the self-attention neural networks to mine the deep relations and retrieve the cipher-images. In our retrieval model, we design a self-attention multi-layer perceptron module, called SAMLP, to effectively learn global dependencies within representations of cipher-images. Extensive experiments present our encryption algorithm is compression-friendly, ensures no information leakage, and HBSAN significantly outperforms other state-of-the-art models in retrieval performance.