ERNet: Efficient Revealing of Perceptually Encrypted Images With Long-Range Dependency

Abstract

Perceptual encryption (PE) conceals sensitive visual content to protect image privacy, but it also raises the security risk of unauthorized visual content revealing. In particular, position-altering PE schemes jointly introduce local degradation and cross-region pixel rearrangements, resulting in complex local-region short-range and cross-region long-range pixel dependencies. Effective revealing requires the joint modeling of both types of dependencies. However, most existing learning-based revealing methods rely heavily on local modeling assumptions and are mainly effective for low-resolution perceptually encrypted images (PEIs). When extended to high-resolution settings, they often suffer from unsatisfactory revealing fidelity and prohibitively high computational cost. To address these challenges, we propose an efficient revealing network (ERNet) for high-resolution PEIs. Specifically, ERNet first maps the high-resolution images into compact visual semantic features through a cross-scale low-dimensional compressed representation (CLCR) module, which reduces the cost of subsequent global dependency modeling. Then, we design the lightweight long-range dependency modeling (LLDM) module, which consists of receptive field enhanced attention (RFEA) and channel-wise feature hierarchical distillation (CFHD). RFEA combines local and sparse attention to jointly capture short-range and long-range pixel dependencies, while CFHD performs hierarchical channel distillation to select and refine the most informative features. Finally, we introduce a progressive deconvolution high-resolution reconstruction (PDHR) module to progressively recover image structure and fine details. Extensive experiments demonstrate that the proposed ERNet achieves high-fidelity and efficient PEI revealing across different PE schemes, encryption strengths, and resolution scenarios.