Transfer Learning Framework Enhances Digital Image Forgery Detection
A new study published on arXiv introduces a robust framework for detecting digital image forgery using transfer learning and deep convolutional neural networks (CNNs). Addressing the rising challenge of manipulated content due to advanced editing tools, the research integrates compression-aware feature enhancement with hybrid input representations. This approach combines standard RGB images with compression difference-based features (FDIFF) to highlight subtle manipulation artifacts often missed by traditional methods. The framework also employs an adaptive threshold optimization strategy based on the Youden Index to balance true and false positive rates effectively. Experiments conducted on the CASIA v2.0 dataset evaluated multiple pretrained CNN architectures, including DenseNet121, VGG16, and ResNet50. Results indicated that while DenseNet121 achieved the highest accuracy and AUC, ResNet50 provided the most reliable predictions with the highest Matthews correlation coefficient. The findings emphasize that accuracy alone is insufficient for forensic applications, highlighting the need to minimize false negatives. This framework significantly improves artifact visibility and classification robustness, offering a viable solution for real-world digital forensics and information security scenarios.
Wire timeline
Transfer Learning Framework Enhances Digital Image Forgery Detection
A new study published on arXiv introduces a robust framework for detecting digital image forgery using transfer learning and deep convolutional neural networks (CNNs). Addressing the rising challenge of manipulated content due to advanced editing tools, the research integrates compression-aware feature enhancement with hybrid input representations. This approach combines standard RGB images with compression difference-based features (FDIFF) to highlight subtle manipulation artifacts often missed by traditional methods. The framework also employs an adaptive threshold optimization strategy based on the Youden Index to balance true and false positive rates effectively. Experiments conducted on the CASIA v2.0 dataset evaluated multiple pretrained CNN architectures, including DenseNet121, VGG16, and ResNet50. Results indicated that while DenseNet121 achieved the highest accuracy and AUC, ResNet50 provided the most reliable predictions with the highest Matthews correlation coefficient. The findings emphasize that accuracy alone is insufficient for forensic applications, highlighting the need to minimize false negatives. This framework significantly improves artifact visibility and classification robustness, offering a viable solution for real-world digital forensics and information security scenarios.
cs.AI updates on arXiv.org