Masked image: Visually protected image dataset privacy-preserving scheme for convolutional neural networks

With the widespread use of the internet and advancements in computing and data storage technologies, large amounts of data can be collected and analyzed to explore hidden knowledge and patterns in various areas. Among different data types, images have become an increasingly important information carrier, especially in computer vision research. However, the image data often contains valuable sensitive information of users, such as personal identifiers and biometric information. During the transmission process, once adversaries obtain the image data, it may lead to severe consequences. To solve the image data leakage problem, a lot of image privacy-preserving schemes are proposed. Nevertheless, in most existing schemes, the utility of an image will also be broken while ensuring data privacy. Meanwhile, how to balance privacy and utility is always a nerve-wracking challenge in image privacy protection. In this paper, we focus on designing a privacy-preserving scheme that protects the visual content of an image while retaining its availability for convolutional neural networks (CNN) training and prediction. Specifically, we first use the edge detection technique to discover original image features. Then, we carefully design a noise generation method in which the generated noises can guarantee distance-based indistinguishability while minimally affecting image features. Therefore, with the proposed scheme, the visual content of an image can be protected, and it can still be used for CNN training and prediction. Moreover, we empirically evaluate our proposed scheme with various evaluation criteria. The results demonstrate that the visual content of an image is indeed protected while the accuracy of the trained CNN model is available.