Encrypted Image Visual Security Index via Non-Local Recognizable Degree Evaluation

Abstract

With the development of perceptual image encryption techniques, visual security evaluations of perceptually encrypted images have gained much attention. Current visual security indices isolate neither global nor local visual security evaluations. They ignore the fact that humans can reorganize partial local information to infer global information and local information can possibly be leaked in any position of the encrypted image. To address this problem, we propose a block-based non-local recognizable degree measure with a global structure similarity measure as a visual security index. In our index, the non-local searching strategy is utilized to capture leaked local information in any position of an encrypted image. The recognizable degree is evaluated by appearance recognizability and spatial structural recognizability involving human visual properties. In addition, weighted Minkowski pooling is adopted to evaluate the overall recognizable degree of all blocks depending on highly recognizable blocks. Furthermore, this overall recognizable degree is adjusted by global structure similarity, which is used to alleviate the global region structure distortion problem induced by the block partition. Experimental results demonstrate the sharpness and good robustness of our proposed index on different databases and various encryption types.