Forensic Prompting with Dual-Action Policy Optimization for Vision-Language Forgery Detection and Localization

Image forgery is rapidly evolving, rendering forensic traces increasingly subtle and readily attenuated by post-processing. Although vision-language prompting can inject priors, open-ended LLM-generated prompts are difficult to constrain, and naive language descriptions can introduce semantic perturbations. To address these challenges, we propose Forensic Prompting with Dual-Action policy optimization (FPDA) for vision-language forgery detection and localization, where the Forensic Prompting Module (FPM) constructs a structured and reproducible forensic prompt bank and supports optional text input as a reliability-aware cue for stable conditioning. Moreover, Dual-Action Policy Optimization (DAPO) is applied to learn sample-adaptive evidence usage by routing forensic prompts and scheduling localization refinement on a per-image basis, stabilizing discriminative cues and improving mask spatial consistency. Extensive experiments are conducted on multiple public datasets covering manual manipulations, diffusion content, face forgeries, and text-enabled settings, demonstrating favorable detection and localization performance over representative state-of-the-art methods under comparable evaluation protocols.

Modern image editing and generation tools are making forged images harder to recognize. The traces left by manipulation can be very small, and common operations such as compression or resizing can make them even harder to see. This paper studies how to decide whether an image has been manipulated and how to find the changed regions. We propose a method that organizes reliable signs of image manipulation in a structured and reproducible way, such as unusual compression patterns, unnatural boundaries, and inconsistent textures. Because different forged images may show different signs, the method also learns which signs to focus on and how to refine the suspected regions for each image. It can use image descriptions when they are available, while reducing the influence of descriptions that are missing, noisy, or unreliable. Across several public datasets, our method performs well in both detecting forged images and locating manipulated regions. This can support content authentication, misinformation analysis, and digital integrity workflows.