The rapid proliferation of deepfake technology threatens information integrity. We propose an Auxiliary-LossGuided Feature Learning framework addressing generalization and class imbalance issues in existing detectionmethods. Using a Swin Transformer b...
The rapid proliferation of deepfake technology threatens information integrity. We propose an Auxiliary-LossGuided Feature Learning framework addressing generalization and class imbalance issues in existing detectionmethods. Using a Swin Transformer backbone with three lightweight auxiliary encoders learning forgery, content,and common features separately, the key innovation guides the main network through a comprehensive losssystem rather than direct feature ensembling. This enhances feature quality without inference overhead. DFDCexperiments achieved 99.7% accuracy, 0.9999 AUC, and notably 99.4% recall on real samples (minority class),demonstrating effectiveness in class-imbalanced environments