Hybrid framework for image forgery detection and robustness against adversarial attacks using vision transformer and...

Area of Science:

• Computer Science
• Artificial Intelligence
• Cybersecurity

Background:

• Digital media is increasingly used as evidence, but image manipulation poses significant risks.
• The prevalence of image editing tools necessitates robust methods for detecting fraudulent visual content.
• Concerns about manipulated photos and videos impact individuals and society.

Purpose of the Study:

• To develop and evaluate a deep learning framework for distinguishing authentic images from manipulated ones (copy-move and splicing).
• To enhance model resilience against adversarial attacks through adversarial training.
• To assess the proposed method's performance on multiple established image forensics benchmarks.

Main Methods:

• Utilized a pre-trained Vision Transformer (ViT) for feature extraction.
• Employed a Support Vector Machine (SVM) for binary classification of images.
• Implemented adversarial training to improve robustness against manipulation.
• Evaluated the model on CASIA v1.0, CASIA v2.0, MICC-F220, MICC-F2000, and MICC-F600 datasets.

Main Results:

• The combined ViT-SVM approach demonstrated strong performance in differentiating authentic from manipulated images.
• Adversarial training significantly enhanced the model's robustness against sophisticated image forgeries.
• The proposed framework achieved competitive results compared to existing image manipulation detection methods.
• Extensive validation across multiple benchmarks confirmed the methodology's potential.

Conclusions:

• The developed deep learning framework offers a promising solution for detecting image forgeries.
• The approach provides improved robustness and accuracy in identifying manipulated visual content.
• This research contributes to enhancing the reliability of digital evidence in the face of evolving manipulation techniques.