A privacy-preserving content-based image retrieval method based on deep learning in cloud computing. (1st October 2022)

• Record Type: Journal Article
• Authors: Ma, Wentao; Zhou, Tongqing; Qin, Jiaohua; Xiang, Xuyu; Tan, Yun; Cai, Zhiping
• Is Part Of: Expert systems with applications. Volume 203(2022)
• Publication Date: 2022-10-01
• Subjects: Image retrieval -- Privacy-preserving -- Deep convolutional network -- Edge computing -- CBIR; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2022.117508
• Languages: English
• ISSNs: 0957-4174

Abstract

Abstract: Privacy-preserving Content-Based Image Retrieval (CBIR) method is a promising technology to achieve data confidentiality and searchability in cloud-assisted multimedia (i.e., image or video) data environment. However, inappropriate feature-preserving mechanisms and inefficient ciphertext descriptors resulted in lower performance than expected. Therefore, how to design encryption techniques with high security and how to extract effective features from ciphertext images still hinder privacy-preserving CBIR. For this goal, we propose a privacy-preserving image retrieval based on deep convolutional network features. First, a novel hybrid encryption technique is designed to encrypt images and an improved DenseNet model is fine-tuned by using the encrypted images to construct a feature extractor. The encrypted images and fine-tuning feature extractor are then uploaded to cloud server. Meanwhile, secure CBIR service is executed in the cloud server. We conduct experiments on two public benchmark datasets for performance evaluation in terms of mAP and accuracy. As demonstrated in the experimental results, the proposed method can achieve superior result compared with the existing methods, improving the performance on the two metrics by relatively 1.9% and 10%, respectively. Furthermore, the computational cost and parameters of depthwise separable convolution adopted by the improved DenseNet model are 8 to 9 times smaller than that of standard convolutions of the original DenseNet at only a small reduction in accuracy. Highlights: Preserving color features by replacing RGB channels. The texture features are preserved by scrambling the pixel bit matrix. A hybrid privacy-preserving mechanism for color and texture information. The features of encrypted images are extracted by CNN on cloud server. Experimental results show that our method is highly competitive.