Subjective and Objective Evaluation of Visual Security in Perceptually Encrypted Images

Research onvisual security evaluation (VSE) of perceptually encrypted images largely depends on perceptually encrypted datasets with subjective annotations. However, existing datasets are constrained by limited scene diversity, a narrow range of encryption algorithms, and small scales, restricting the advancement of VSE research. To address these issues, this work investigates both subjective and objective VSE for perceptu ally encrypted images. A large-scale and comprehensive benchmark, named Color Encrypted Image Dataset (CoEIDs), is first constructed. CoEIDs comprises 36 color reference images encrypted by 10 independent and 3 hybrid encryption algorithms under 5-10 strength levels, yielding 2,880 encrypted samples. Standardized subjective testing procedures are employed to obtain visual security (VS) and visual quality (VQ) scores for each image. Based on CoEIDs, we further propose a full-reference VSE model that jointly exploits Global Similarity and Local Residual features (GSLR) to effectively capture both structural consistency and perceptual variations caused by encryption. Extensive experiments on CoEIDs and two public datasets (IVC and PEID) demonstrate that GSLR achieves state-of-the-art performance, which further verifies both the reliability of CoEIDs as a benchmark and the effectiveness and generalization ability of the proposed GSLR framework. The dataset and corresponding algorithms will be publicly available at: https://github.com/Xiaodong-Bi/CoEIDs.