Aggregating global-scale pixel-wise forgery cues within a graph

Area of Science:

• Computer Vision
• Digital Forensics
• Machine Learning

Background:

• Deep image inpainting creates realistic forgeries, challenging traditional detection methods due to local coherence and semantic consistency.
• Existing detectors struggle with seamless forgeries, necessitating advanced techniques for accurate localization.

Purpose of the Study:

• To propose a novel Fine-Grained Graph Convolution Network (IFL-GCN) for effective inpainting forgery localization.
• To enhance the sensitivity and robustness of forgery detection against high-fidelity and diverse inpainting artifacts.

Main Methods:

• Introduced a pixel-wise graph construction for direct integration of local forgery traces across the entire image.
• Developed a Fidelity-aware Weighted Loss (FW loss) to calibrate learning objectives based on forged content fidelity.
• Proposed Forgery Intensity Mixup (FIM) augmentation to improve generalization across diverse inpainting artifacts.

Main Results:

• IFL-GCN achieves state-of-the-art performance on 6 mainstream forgery detection benchmarks.
• Outperformed the closest competing method by 6.7% in average F1 score across all inpainting forgery test sets.
• Demonstrated enhanced sensitivity to subtle, high-fidelity forgeries and improved robustness.

Conclusions:

• The proposed IFL-GCN effectively addresses the challenges posed by deep inpainting forgeries.
• The pixel-wise graph, FW loss, and FIM augmentation contribute to superior inpainting forgery localization.
• IFL-GCN represents a significant advancement in digital forensics for detecting manipulated images.