Abstract
Traditional deepfake detection methods typically treat the problem as a binary classification task. However, this approach often fails to account for the significant impact of image quality on visual artifacts. This paper introduces DeepFidelity, a novel framework that redefines deepfake detection as a continuous forgery fidelity regression problem. By incorporating image quality assessment and a specialized Transformer architecture (SSAAFormer) that exploits facial symmetry, the proposed method achieves state-of-the-art performance across multiple benchmarks, including a perfect 100% accuracy on the Celeb-DF dataset.
Rethinking the Detection Paradigm
Most existing deepfake detectors are trained to output a simple binary label: 0 for fake and 1 for real. While intuitive, this binary approach ignores the underlying distribution of image quality. In real-world scenarios, a low-quality real image—perhaps blurry or poorly lit—may share more visual characteristics with a deepfake than with a high-quality real image. Conversely, a high-fidelity deepfake might appear "more real" to a standard classifier than a grainy CCTV frame of a genuine person.
This overlap in the quality distribution leads to high error rates and poor generalization. DeepFidelity addresses this by mining the "perceptual forgery fidelity" of face images, treating the "realness" of a face as a continuous spectrum influenced by its visual quality rather than a discrete state.
The DeepFidelity Framework
The DeepFidelity framework consists of two primary innovations designed to bridge the gap between quality and authenticity:
1. Perceptual Forgery Fidelity Assessment
Instead of binary labels, the model is trained to predict a fidelity score between 0.0 and 1.0. Using an Image Quality Assessment (IQA) tool, images are mapped into specific ranges:
- Real faces are mapped to scores between 0.6 and 1.0, where higher quality corresponds to a higher score.
- Fake faces are mapped to scores between 0.0 and 0.4. While higher quality fakes receive higher scores within this bracket, they remain distinct from the real distribution.
This mapping creates a "buffer zone" (0.4 to 0.6) and forces the model to learn the nuances of how quality affects forgery traces, preventing the detector from misclassifying low-quality real images as fakes.
2. SSAAFormer Architecture
To extract robust features for this regression task, the authors developed the Symmetric Spatial Attention Augmentation based vision Transformer (SSAAFormer). Standard Vision Transformers (ViTs) often struggle with local feature extraction in shallow layers. SSAAFormer overcomes this by exploiting the natural symmetry of human faces.
In the early stages of the network, the model performs a "flipping" operation on feature maps and weights them together. This Symmetric Spatial Attention allows the model to capture long-distance geographic relationships—such as comparing the left eye to the right eye—early in the processing pipeline. This makes the feature representation much more sensitive to structural inconsistencies common in deepfakes, such as mismatched reflections or asymmetrical features.
Experimental Evaluation
The model was evaluated on three major benchmarks: FaceForensics++ (FF++), Celeb-DF (v2), and WildDeepfake. The results demonstrate the superiority of the fidelity-based approach:
- Celeb-DF (v2): DeepFidelity achieved a perfect 100% Accuracy and AUC, significantly outperforming previous state-of-the-art models like RECCE and SFDG. This is particularly notable as Celeb-DF is considered a "gold standard" for testing high-fidelity detection.
- WildDeepfake: On this challenging dataset collected from the internet, the model reached 88.34% Accuracy, a nearly 4% improvement over previous methods.
- Cross-Dataset Generalization: When trained on FF++ and tested on the unseen WildDeepfake dataset, it achieved an AUC of 71.60%, proving that the fidelity-based approach generalizes better to unknown forgery techniques than traditional binary classifiers.
Conclusion
By shifting the focus from binary classification to a quality-aware fidelity spectrum, DeepFidelity provides a more robust and equitable approach to deepfake detection. The inclusion of facial symmetry as a structural prior in the SSAAFormer architecture further enhances the model's ability to detect sophisticated forgeries. Future work will focus on developing IQA tools specifically tailored for deepfake detection and exploring more advanced regression architectures to further refine the fidelity mapping.
Summary updated