Privacy Preserving Image Matching Scheme with Aggregated Local Descriptors

Detecting similar images, without violating their privacy, is a desirable method for many real-world applications. Existing schemes are based either on global features that are too rigid to describe images or local features which achieve more accurate results with higher complexity. The current local feature aggregation methods rely on constructing and sharing a visual word vocabulary, which leads to serious privacy concerns. In this paper, we proposed efficient schemes, of different security guarantees, for matching private images while maintaining their privacy. To do so, we developed a new method for aggregating the local features, by utilizing the locality-sensitive hashing functions without the need for predefined vocabulary. Extensive experiments on real-world datasets demonstrate the practical usage and security of our schemes.