Design of an Improved Method for Forgery Detection using DenseNet, Haralick Features, and EfficientNet-B3 with Adversarial Fine-tuning

Abstract

The sudden appearance of high-quality image and video forgeries, such as deepfakes and splicing, has urgently called for more  advanced and generalizable detection frameworks. Most existing forgery detection methods suffer from limited robustness and generalization across different forgery techniques and modalities. To address these limitations, we extend a unified multimodal image and video forgery detection framework by using improved feature extraction and fusion techniques. For image forgery detection, our framework combines the strengths of a DenseNet-based deep feature extraction technique with Haralick texture features to capture both spatial and texture-based manipulations. DenseNet is selected because, by using dense connections, it can reuse features in a very effective manner; hence, it provides a strong mechanism for detecting even fine-grained forgeries. It incorporates Haralick features into its architecture so that any texture anomalies arising from manipulations such as copy-move sets can be identified. The combination of these features is achieved via an attention-based mechanism that dynamically balances the contributions of both feature types based on the nature of the forgery. We also use a pre-trained EfficientNet-B3, fine-tuned with GAN-generated adversarial examples, to make our model more robust to sophisticated forgeries. In the video forgery detection framework, 3D ResNet is incorporated for spatiotemporal feature extraction, LSTM for capturing long-term temporal dependencies, and Temporal Convolutional Networks for ensuring shortterm temporal consistency. A dual attention mechanism is utilized to emphasize manipulated spatial regions and key temporal intervals, thereby improving the accuracy of video forgery detection. It achieved competitive accuracy-95-97% on images and 92-95% for videos-along with improved adversarial robustness, while at the same time presenting a scalable solution for practical forgery detection across different domains.