MS-FPR: Optimized Multi-Scale Training Strategy for Reducing False Positives in Deepfake Detection

The deepfake technique has made significant progress but also brings potential risks to the spread of misinformation. In recent years, various deepfake detection methods have been proposed and have achieved great improvements in accuracy and robustness. However, reducing false alarms, a crucial requirement for real-world applications, has not received sufficient attention. Most detection methods widely use AUC(Area Under Curve) as an indicator to measure the overall performance of the algorithm. Although AUC provides a comprehensive evaluation of the classifier's performance, it does not explicitly reflect misclassification errors, particularly high-confidence misclassified real samples, which are often overlooked. These errors are challenging to rectify by merely adjusting the detection threshold, as this typically results in a trade-off between reducing the false positive rate and preserving recall. To address the problem, we propose a novel multi-scale training strategy and a plug-and-play lightweight two-branch framework to enhance the existing CNN-based deepfake detectors and improve the decision boundary of the model. The framework adopted a novel training strategy that used the original resolution of samples for training. The experimental results demonstrate that our framework can significantly improve the reliability of deepfake detection systems in actual deployment scenarios.