Deployment-Oriented Unified Evaluation Framework for Secure Content-Based Image Retrieval

Secure Content-Based Image Retrieval (SCBIR) schemes successfully safeguard outsourced visual archives through cryptographic transformations while preserving searchability. However, existing evaluation paradigms create a severe performance bottleneck by rigidly treating unencrypted plaintext accuracy as a mandatory, single-metric criterion to judge ultimate system efficacy. This isolated optimization model ignores the underlying secure visual processing trilemma, failing to anticipate hidden adversarial vulnerabilities or computational processing overheads. To address this system deployment limitation, this paper introduces a formal Security–Utility–Efficiency (S-U-E) evaluation framework executing as an integrated multi-criteria decision-making model for cloud-assisted architectures. By projecting candidate encryption schemes onto a standardized three-dimensional valuation space, the framework maps conflicting operational requirements into a unified performance vector. To optimize selection parameters without neutralizing secondary system constraints, a strictly dominant residual sweep operator is introduced to bound primary objective weights while distributing residual mass asymmetrically across secondary goals. We validate our framework using five distinct block-based perceptual encryption (BPE) schemes paired with histogram-based descriptors across diverse natural and clinical image datasets. Through empirical analysis across the three datasets, the framework’s scenario-specific selections reduce adversarial reconstruction leakage metrics by 52.08% to 100.00% and decrease end-to-end execution latencies by 46.60% to 89.79%, while bounding the corresponding utility drops between 0.00% and 3.62%.