New transfer learning framework detects image forgeries with 98% accuracy

A new study published on arXiv (Paper 2605.08167) presents a transfer learning-based approach to digital image forgery detection. The framework introduces a hybrid input representation that fuses standard RGB images with compression difference features (FDIFF), designed to expose subtle manipulation artifacts often missed by traditional methods. By leveraging compression-aware feature enhancement, the system improves the visibility of tampered regions. Additionally, the authors apply a model-specific adaptive threshold optimization using the Youden Index to balance true positive and false positive rates, critical for forensic applications where false negatives can have serious consequences.

Extensive experiments on the CASIA v2.0 dataset, a benchmark for image forensics, evaluated six pretrained CNN architectures: DenseNet121, VGG16, ResNet50, EfficientNetB0, MobileNet, and InceptionV3. Performance metrics included accuracy, precision, recall, F1-score, Matthews correlation coefficient (MCC), and AUC. DenseNet121 achieved the highest accuracy and AUC, while ResNet50 provided the most balanced and reliable predictions with the highest MCC. The study emphasizes that accuracy alone is insufficient in forensics—minimizing false negatives is paramount. Overall, the framework enhances classification robustness and artifact visibility, making it suitable for real-world digital forgery detection.