A New Reinforcement Learning Agent Framework for Digital Image Forgery Detection Using Prioritized Experience Replay

Authors

• Muthana S. Mahdi, Department of Computer Science, College of Science, Mustansiriyah University, Baghdad, IraqFollow
• Saad N. Alsaad, Department of Computer Science, College of Science, Mustansiriyah University, Baghdad, IraqFollow
• Hasanen S. Abdullah, Department of Computer Science, University of Technology, Baghdad, IraqFollow

Abstract

Digital image forgery detection has become an urgent and complex problem in an age when powerful editing tools can easily alter photographs. The familiar maxim ``a picture is worth a thousand words'' can no longer be taken at face value, since even subtle manipulations may conceal or fabricate critical details. Conventional detection methods frequently suffer from highly imbalanced datasets and narrow feature representations that fail to capture the diverse artifacts introduced by modern editing techniques. In response, this work presents a novel framework that casts the forgery detection task as a reinforcement learning problem, enabling an agent to learn a sequence of analysis steps rather than relying on a fixed classification pipeline. This is the first study to deploy a Deep Q-Network (DQN) agent specifically for image forgery detection. The proposed architecture features a dual-branch feature extractor: one branch applies Error Level Analysis to reveal compression inconsistencies. At the same time, the other performs fine-grained noise analysis to detect manipulation traces. These features are supplied to a custom Gym environment designed with balanced sampling to mitigate class imbalance and augmented by a prioritized experience replay mechanism that biases learning toward the agent's most challenging examples. Extensive experiments across multiple datasets, including testing on the latest image manipulation model, Gemini 2.0 Flash, show impressive accuracy rates between 97% and 98%. These results highlight the robustness and efficiency of the proposed approach, offering a promising new direction for digital forensic investigations.

Keywords

Balanced gym environment, Deep q-network, Deepfake, Dual-branch feature extraction, Prioritized experience replay, Reinforcement learning agent, Splice image forgery detection

Subject Area

Computer Science

Article Type

Article

First Page

1316

Last Page

1332

Creative Commons License

This work is licensed under a Creative Commons Attribution 4.0 International License.

How to Cite this Article

Mahdi, Muthana S.; Alsaad, Saad N.; and Abdullah, Hasanen S. (2026) "A New Reinforcement Learning Agent Framework for Digital Image Forgery Detection Using Prioritized Experience Replay," Baghdad Science Journal: Vol. 23: Iss. 4, Article 14.
DOI: https://doi.org/10.21123/2411-7986.5270