Multi-domain Multi-scale DeepFake Detection for Generalization

As deepfake detection technology progresses and application scenarios broaden, it becomes challenging for detection models within a single domain or single scale to handle the complex application environments, leading to limited accuracy and generalization capabilities of the detection models. To address the above problems, we propose a Multi-domain Multi-scale DeepFake Detection model. First, we employ the RetinaFace framework to preprocess and augment the image data, which directs the model’s attention to the relevant regions of the image. Next, we utilize the EfficientNet-B0 network for rapid coarse-grained feature extraction. Building on these coarse-grained features, we design two modules to capture forgery traces at multiple levels: a spatial-domain fine-grained feature extraction module based on multi-scale analysis and a frequency-domain fine-grained feature extraction module leveraging the Fast Fourier Transform. Finally, the dual-domain features are fused using a multi-channel attention mechanism, enabling the model to comprehensively analyze images from various perspectives. Additionally, we implement a Particle Swarm Optimization (PSO) strategy to optimize the hyper-parameters of the network model. The proposed model demonstrates significant advantages over other algorithms, particularly in cross-dataset detection, exhibiting strong generalization, robustness, and adaptability. When trained on the FF++ dataset and tested on the Celeb-DF dataset, it achieves an AUC of 91.7%. This represents a 14.8% improvement in cross-dataset performance compared to the best existing model.