Privacy-Preserving TPE-Based JPEG Image Retrieval in Cloud-Assisted Internet of Things

Yakun Ma,Xiuli Chai,Zhihua Gan,Yushu Zhang
DOI: https://doi.org/10.1109/jiot.2023.3301042
IF: 10.6
2024-01-01
IEEE Internet of Things Journal
Abstract:With the large-scale deployment of the Internet of Things, lots of images are generated and outsourced to the cloud to alleviate storage burdens. Encrypted image retrieval has been widely studied as a promising technique for balancing privacy and usability. However, existing schemes have primarily focused on retrieval usability but neglected visual usability, resulting in limitations in image management, even in scenarios where preserved visual information poses no hazard to privacy. In this paper, we are thus inspired to propose a privacy-preserving JPEG image retrieval scheme that effectively enhances retrieval efficiency and accuracy while ensuring low file expansion and excellent thumbnail-preserving accuracy for ciphertext images. Specifically, the well-designed thumbnail-preserving encryption (TPE) enables accurate thumbnail-preserving and high-quality decryption of encrypted images by integrating Huffman coding and information embedding techniques. Moreover, the proposed TPE method fully considers JPEG compression to decrease the ciphertext file expansion and gain good format compatibility. Additionally, adaptive encryption key generation is designed to minimize key storage and considerably bolster security. Also, the cloud can generate preview thumbnails for uploaded TPE-encrypted images, and then extract the Hue-Saturation-Value (HSV) and uniform Local Binary Pattern (ULBP) features from thumbnails instead of encrypted images to boost retrieval efficiency and accuracy. Experimental results show that the peak signal-to-noise ratio (PSNR) of thumbnail-preserving accuracy and decrypted image quality reaches 52dB and 61dB, respectively, the file expansion is minimally limited to 6%, and the mean average precision (mAP) of retrieval attains 64%, indicating that our scheme significantly outperforms the state-of-the-art ones.
What problem does this paper attempt to address?