CAL-Net: a continual adaptive learning network for deepfake detection using multi-track fusion and domain adaptation

Abstract The fast advancement of deepfake generation methods poses a serious challenge to the detection of deepfakes using traditional methods. These methods rely on static models and low-level artifact analysis, limiting their ability to generalize across evolving manipulations. This widening gap requires a new approach to deepfake detection methods that are both adaptive and dynamic. To address this problem, we introduce the Continual Adaptive Learning Network (CAL-Net), a novel method that combines multi-track feature extraction and dynamic domain adaptation for deepfake detection. CAL-Net consists of three complementary feature extractors: frequency domain analysis to identify inconsistencies introduced by deepfake manipulations, Data-Efficient Image Transformer (DEIT) for representing global information, and Xception network for learning local information. By combining these multi-track features, CAL-Net effectively detects important manipulation traces. We also propose a domain adaptation module that leverages cosine similarity for effective domain detection. This module dynamically adapts to new manipulation methods while mitigating the issue of catastrophic forgetting. To the best of our knowledge, this is the first work to combine multi-track feature extraction and domain adaptation with continual learning for deep fake detection. Evaluating on the Continual Deepfake Detection Benchmark (CDDB) dataset, CAL-Net achieved 94% accuracy and outperformed existing methods. The proposed approach demonstrates cross-domain generalization and robustness to diverse and new deepfake attacks. The research emphasizes the need for continual learning and multi-track feature fusion to make deepfake detection systems scalable.