Audio-visual deepfake detectors are often developed under a single dataset or manipulation regime, but real-world training collections assembled from multiple sources are more complicated. The new arXiv preprint argues that multi-source training data can contain heterogeneous attacks and uneven source distributions, making it harder for a detector to learn robust cues.
The paper introduces shared-private cross-modal discrepancy learning as a response. The approach appears to model both shared and private information across audio and visual modalities, and to account for imbalance across sources. The abstract begins to contrast this with training separate detectors, but the provided text cuts off before the comparison is completed.
Because only one source was available, there are no independent results to compare. The contribution is positioned as a way to improve audio-visual deepfake detection when training data is multi-source and imbalanced.