Digital Image Forgery Detection Using Transfer Learning

Researchers present a transfer learning framework for detecting digitally forged images by combining RGB data with compression-difference features and optimized thresholds. Testing across multiple CNN architectures on the CASIA v2.0 dataset shows DenseNet121 achieves highest accuracy while ResNet50 provides most reliable predictions, addressing critical forensic security needs.

Digital image manipulation has become increasingly sophisticated with accessible editing tools, creating substantial challenges for forensic authentication and content verification across industries from media to legal proceedings. This research tackles the problem through a technical framework that enhances traditional CNN approaches with compression-aware feature analysis, designed to reveal subtle manipulation artifacts invisible to standard detection methods.

The hybrid input strategy combining RGB imagery with FDIFF (compression difference-based features) represents a meaningful advancement in forensic detection capabilities. Rather than relying solely on standard image data, the approach explicitly targets the compression artifacts left behind by editing processes. The integration of Youden Index-based adaptive thresholds further refines detection by optimizing the balance between true positives and false negatives—a critical consideration in forensic applications where missing a forgery carries significant consequences.

The comparative evaluation across six major architectures (DenseNet121, VGG16, ResNet50, EfficientNetB0, MobileNet, InceptionV3) reveals important nuances about model selection in forensic contexts. While DenseNet121 shows superior accuracy and AUC scores, ResNet50's highest Matthews correlation coefficient suggests better overall reliability for real-world deployment. This distinction underscores that accuracy alone provides an incomplete picture for security applications.

For enterprises managing content authentication—news organizations, government agencies, and legal institutions—improved forgery detection directly reduces fraud risks and operational vulnerabilities. The framework's demonstrated robustness on CASIA v2.0 suggests practical applicability to production systems. Future development should focus on testing against emerging deepfake technologies and establishing performance benchmarks on contemporaneous manipulation techniques that continue evolving beyond the dataset's scope.

• →Transfer learning framework combining RGB images with compression-difference features improves detection of subtle image manipulation artifacts.
• →DenseNet121 achieves highest accuracy and AUC, while ResNet50 provides most balanced predictions with highest Matthews correlation coefficient.
• →Youden Index-based adaptive threshold optimization significantly improves classification reliability by balancing true positives and false positives.
• →Framework tested on CASIA v2.0 dataset demonstrates robustness across multiple CNN architectures in forensic applications.
• →Results emphasize that accuracy alone is insufficient for forensics; minimizing false negatives is critical for real-world deployment.