Pixel-Inconsistency Modeling for Image Manipulation Localization.

Area of Science:

• Computer Vision
• Digital Image Forensics
• Machine Learning

Background:

• Deep neural networks have advanced image forgery localization but struggle with generalization to new datasets and robustness against image perturbations.
• Existing methods often fail in real-world scenarios due to limitations in handling unseen data and minor image alterations.

Purpose of the Study:

• To develop a generalized and robust image manipulation localization model.
• To address the lack of generalization and robustness in current deep learning-based forgery detection techniques.
• To enhance image integrity through reliable manipulation detection.

Main Methods:

• The model analyzes pixel inconsistency artifacts, leveraging the pixel correlations introduced by the demosaicing process in image signal processors.
• It employs masked self-attention mechanisms to model global pixel dependencies and a local pixel dependency stream to identify manipulation clues.
• Novel Learning-to-Weight Modules (LWM) integrate features from both streams, and a Pixel-Inconsistency Data Augmentation (PIDA) strategy focuses on inherent pixel-level artifacts.

Main Results:

• The proposed method successfully extracts inherent pixel-inconsistency forgery fingerprints.
• Achieved state-of-the-art generalization and robustness in image manipulation localization tasks.
• Demonstrated superior performance compared to 16 representative detection models across 12 diverse datasets.

Conclusions:

• The developed model provides a generalized and robust solution for image manipulation localization by focusing on pixel-level inconsistencies.
• The approach effectively overcomes the limitations of existing methods in real-world applications.
• This work sets a new benchmark for robust and generalizable image forensics.