M²RL-Net: Multi-View and Multi-Level Relation Learning Network for Weakly-Supervised Image Forgery Detection

Jiafeng Li, Ying Wen, Lianghua HeAAAI 2025 · DOI

Abstract

Image forgery detection has traditionally relied on fully supervised models requiring labor-intensive pixel-level masks. We present M²RL-Net, a framework designed for weakly-supervised image forgery detection (W-IFD) that utilizes only image-level labels. By leveraging multi-view feature representations and multi-level relation learning, M²RL-Net achieves state-of-the-art performance in both detection and localization, often surpassing fully supervised counterparts.

Introduction

The proliferation of sophisticated image manipulation tools necessitates robust forgery detection. While fully supervised methods achieve high accuracy, their reliance on pixel-level annotations limits scalability. Weakly-supervised methods offer a more scalable alternative but typically struggle with precise localization. M²RL-Net addresses these challenges by modeling internal image relationships and contrasting features across different views.

Methodology

M²RL-Net consists of three primary modules designed to extract and refine forgery-related features without explicit mask guidance.

The M²RL-Net architecture integrates Multi-View Feature Representation, Patch-level Self-Consistency Learning, and Feature-level Contrastive Learning.

Multi-View Feature Representation (MFR)

The network employs a dual-stream architecture. The RGB stream captures visual artifacts like unnatural edges, while the Noise stream identifies inconsistencies in noise distributions through a learnable constrained convolution layer. This multi-view approach ensures the model captures both semantic and statistical anomalies.

Patch-level Self-Consistency Learning (PSL)

To enable localization under weak supervision, PSL analyzes patch-to-patch attention maps from the Transformer encoder. By recalibrating these maps using 1D convolutions, the model identifies inconsistent patches, effectively generating pseudo-localization regions that guide the learning process without human-drawn masks.

Feature-level Contrastive Learning (FCL)

FCL enhances feature discriminability by building prototypes for authentic and tampered classes. It performs self-contrast within individual views and multi-view contrast between RGB and Noise streams. An Exponential Moving Average (EMA) strategy is used to update these prototypes, ensuring stability against noisy pseudo-labels.

Experimental Results

The model was trained on the CASIAv2 dataset and evaluated across five unseen datasets (CASIAv1, Columbia, COVER, NIST16, and IMD20) to test generalization.

Performance comparison across multiple datasets shows M²RL-Net outperforming both unsupervised and weakly-supervised baselines.

Detection and Localization Performance

M²RL-Net demonstrated significant improvements over existing methods. It achieved a 42.7% improvement in average Image-level F1 score compared to the supervised MVSS-Net. In cross-dataset evaluations, it outperformed the state-of-the-art weakly-supervised method (WSCL) by 19.7% in AUC and 16.3% in Image-F1. Remarkably, its localization performance (P-F1) remained competitive with fully supervised models like ManTra-Net, despite the lack of pixel-level training data.

Generalization and Robustness

The results indicate that modeling internal patch relationships provides a powerful substitute for manual annotations. The high specificity scores suggest that M²RL-Net is particularly effective at minimizing false positives, a common issue in forgery detection. While performance on copy-move datasets like COVER was slightly lower, the overall generalization across diverse tampering types (splicing, inpainting) is robust.

Conclusion

M²RL-Net represents a significant advancement in weakly-supervised image forgery detection. By combining multi-view analysis with patch-level and feature-level relation learning, it bridges the gap between weakly and fully supervised models. Future work will explore extending this framework to language-driven image manipulations.

Summary updated

BibTeX
@inproceedings{li2025m,
  author = {Jiafeng Li and Ying Wen and Lianghua He},
  title = {{M²RL-Net: Multi-View and Multi-Level Relation Learning Network for Weakly-Supervised Image Forgery Detection}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i5.32501},
}
Figure