Title: Divide and Conquer: Reliable Multi-View Evidential Learning for Deepfake Detection

Abstract:

As generative models advance, deepfakes have reached a level of semantic realism that is nearly flawless, with forensic evidence now confined to faint structural irregularities. Traditional single-view detection methods frequently struggle with generalization in this landscape, as strong semantic signals tend to drown out delicate artifact indicators within mixed feature representations. This disparity results in predictions that are highly confident but fragile, a issue we identify as the Semantic Masking Effect.

To overcome these limitations, we introduce a robust framework named Divide-and-Conquer Multi-View Evidential Learning (DiCoME) designed specifically for deepfake detection. The framework operates in two stages. First, during the "Divide" phase, we utilize Geometric View Purification to disentangle the mixed representation space via structured geometric projections. This step mitigates semantic interference within representations sensitive to artifacts, establishing a basis for distinct yet complementary semantic and artifact views that are decorrelated.

In the subsequent "Conquer" phase, the system applies Uncertainty-Aware Evidential Learning to integrate these separate perspectives. By formally capturing the "epistemic conflict" arising between semantic and artifact signals, this approach yields well-calibrated uncertainty estimates rather than imposing strict deterministic outcomes. Comprehensive testing on various benchmarks reveals that our proposed method consistently surpasses current state-of-the-art techniques in terms of generalization capabilities, while also delivering trustworthy deepfake detection through reliable uncertainty quantification. The source code can be accessed at https://github.com/kxl0825/DiCoME.git.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC