Skip to content
Book Open access

SAFER: Sequential Adaptive Forensic Evidence Reasoning for Generalizable Visual Forensics

Nov 2026 · Proceedings of the 34th ACM International Conference on Multimedia · 0 citations · 17 references

Abstract

Image forensics has become increasingly challenging in the era of heterogeneous manipulations, where editing, AI-generated content, and deepfake synthesis expose highly diverse forensic cues across tracks and samples. Existing methods predominantly follow a static one-shot paradigm, in which all inputs are processed with the same computation graph and evidence pathway. Such a design is increasingly restrictive for generalizable forensics, where different samples and forgery tracks often require different types and amounts of evidence. In this work, we propose SAFER (Sequential Adaptive Forensic Evidence Reasoner), a prior-guided forensic agent that reformulates forgery analysis as a budget-aware sequential evidence reasoning problem. Given an input image, SAFER first derives an initial forensic prior using a lightweight multi-scale front-end. Conditioned on this prior, a policy network then adaptively queries one or more track-aware experts from a heterogeneous expert pool under a limited computation budget. The queried evidence is subsequently integrated by a reliability-aware decision aggregation module to produce the final forgery prediction, tampering mask, and explainable report. To make sequential decision-making trainable, we further introduce a progressive policy learning strategy that combines imitation learning and reinforcement learning. Extensive experiments show that SAFER achieves robust performance across diverse scenarios, while providing a more favorable accuracy--efficiency trade-off than conventional static pipelines. Our results suggest that adaptive sequential evidence reasoning offers a practical new paradigm for real-world multimedia forensics.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.