Seam Carving Forgery Detection Through Multi-Perspective Explainable AI.

Area of Science:

• Computer Science
• Artificial Intelligence
• Digital Forensics

Background:

• Deep learning models like CNNs show promise in detecting image manipulations but lack transparency.
• The black-box nature of CNNs hinders trust in critical applications such as digital forensics.
• Content-aware image forgeries, like seam carving, pose a significant challenge for detection.

Purpose of the Study:

• To develop and validate an interpretable framework for detecting image forgeries, specifically seam carving.
• To enhance the trustworthiness of AI systems in digital forensics by providing explainable decisions.
• To address the limitations of current deep learning models in providing transparent explanations for forgery detection.

Main Methods:

• Proposed a novel framework, E-XAI (Ensemble Explainable AI), integrating multiple explainability techniques.
• Combined SHAP (SHapley Additive exPlanations) for pixel-level feature attribution and Grad-CAM for region-level localization.
• Trained a custom CNN and validated the E-XAI framework on a balanced dataset of 10,300 images.

Main Results:

• Achieved high classification performance with 95% accuracy and 99% precision for the forged class on an unseen test set.
• Demonstrated the framework's robustness against JPEG compression, a common real-world perturbation.
• E-XAI provided transparent, visual evidence of how the model identifies subtle forgery artifacts, enhancing interpretability.

Conclusions:

• The E-XAI framework offers a robust end-to-end pipeline for interpretable image forgery detection.
• Integrating explainability techniques improves the reliability and trustworthiness of AI in information security.
• This approach provides crucial insights into the decision-making process of CNNs for image manipulation detection.