REFORGE: A Robust Ensemble for Image Forgery Detection and Localization in Social Network Images

Image forgery detection and localization are critical for verifying the authenticity of digital images. This paper presents REFORGE, an ensemble dual-branch deep network that combines a classification branch for detecting manipulated images and a segmentation branch for localizing tampered regions, enhanced with a reinforcement learning (RL) based refinement module. The classification branch identifies forged images with high precision, while the segmentation branch produces pixel-level tampering masks. A lightweight RL agent then iteratively improves the predicted mask, treating each pixel as an agent to maximize a reward for accurate localization. We integrate outputs from multiple state-of-the-art models in an ensemble fashion to boost generalization and robustness. Extensive experiments on standard benchmarks (CASIA v2 and augmented social-network degraded images) demonstrate that our method achieves near-perfect forgery detection (accuracy ∼98.9%, AUROC ∼0.997) and significantly improved localization on tampered images (mean IoU = 0.570, median = 0.644, Q1/Q3 = 0.323/0.852, IoU≥ 0.5: 63.35%, IoU≥ 0.75: 38.36%). We present ablation studies showing the contribution of each component (ensemble learning, dual-branch architecture, and RL refinement) to overall performance. The proposed approach outperforms baseline models and conventional post-processing (e.g. DenseCRF) in both detection and localization. Our work represents a novel integration of ensemble deep learning and reinforcement learning for image forensics, pushing the state-of-the-art in tampering localization accuracy and reliability.