Fast Test-Time Refinement Enhances Robustness of Learned Image Compression

A recent investigation published on arXiv (2608.15113) delves into test-time refinement (TTR) as a potential defense for learned image compression (LIC) systems. Although LIC exhibits impressive rate-distortion performance under normal conditions, its deep neural networks are susceptible to adversarial threats. Initially intended to enhance performance in benign scenarios, TTR has been suggested as a gray-box defense; however, its theoretical limitations and overhead have hindered widespread use. This research uncovers an Asymmetric Adversarial Trajectory (AAT) characteristic in LIC systems, indicating a quicker transition from adversarial to benign states compared to the opposite. By utilizing this trait, the authors introduce an efficient TTR approach that minimizes iterations and computational demands while ensuring robustness. The study also broadens its evaluation to include white-box scenarios and various attack types, filling previous research gaps. The results establish a theoretical basis for TTR's robustness and highlight its practical applicability for reliable standardized codecs. This paper, authored by a team of researchers, underscores ongoing advancements in AI-driven image compression.