Protecting Consumer Electronics Human-Computer Interactive Verification Security via Anomaly-Aware Reconstruction-Guided Forgery Localization

With the popularity of computer vision, many consumer electronic applications now require users to interactively authenticate their identities through cameras while handling business. However, numerous spoofing technologies were proposed in recent years, which enables attackers to perform manipulation during verification. Such manipulations can easily bypass the human-computer-interaction verification, allowing attackers to impersonate others to commit crimes or fraud. Diffusion, as the SOTA generative model, can flexibly reprint image content according to prompt guidance, evading traditional forgery detection techniques. Therefore, there is an urgent need for countermeasures on diffusion manipulation and protects data security of consumer electronics applications. However, existing studies suffer from overfitting specific forgery traces in the training dataset with limited ability on unseen diffusions. In this paper, we propose a novel diffusion-manipulated forgery localization framework leveraging Anomaly-aware Image Reconstruction Deviation (AIRD). Specifically, we train a reconstruction model capable of recovering detailed texture of authentic images, regarding various spoof traces as uniform anomalies and conducts segmentation according to reconstruction deviation. In this case, the proposed method avoids overfitting specific generators, thus achieving better generalization performance in cross-domain scenarios. Extensive experiments on the multiple datasets showcase our effectiveness in extracting general forgery artifacts facing unseen scenarios and diffusion models.