An audio-visual deepfake detector called DF-MoE was the top reported method in all four MAVOS-DD evaluation setups: closed-set, open-set model, open-set language and open-set full. On closed-set, it reported 0.99 mAP, 0.99 AUC and 97.73% accuracy. The corresponding figures were 0.98, 0.98 and 93.64% for open-set model; 0.99, 0.99 and 94.55% for open-set language; and 0.98, 0.98 and 92.95% for open-set full. The result comes from an arXiv preprint, version 1, dated 24 August 2026.
The AVLips comparison did not produce a clean sweep: DF-MoE reported 0.97 mAP, 0.97 AUC and 93.77% accuracy, while LipFD recorded higher accuracy at 95.27%.
A multimodal design tested across datasets
The paper asks whether diverse high-level audio-visual cues from pretrained models, combined through a sparse Mixture-of-Experts backbone, can reduce dataset-specific overfitting and improve generalization across deepfake generation methods. DF-MoE extracts multiple semantic cues with pretrained models, sends them through specialized adapters and combines them in a sparse Mixture-of-Experts transformer for real-versus-fake classification. In ordinary terms, the system uses several specialized processing paths before making one classification decision.
The evaluation covered five datasets using in-domain, open-set and cross-dataset protocols. A supplemental test used Celeb-DF (v2) in a video-only setting. The reported measures were mean average precision, or mAP, area under the ROC curve, or AUC, and accuracy, the percentage of videos classified correctly.
The reported counts included over 3,000 real and over 4,000 manipulated AVLips videos; 21K MAVOS-DD training videos, 4K validation videos and over 60K test videos; 766 real and 14,472 fake PolyGlotFake videos; 2,010 real and 1,693 fake BioDeepAV videos; and 500 real plus over 19,000 fake FakeAVCeleb videos.
The wider comparisons are less consistent
Cross-dataset testing produced a mixed set of scores. On PolyGlotFake, DF-MoE reported 0.93 mAP, 0.94 AUC and 98.49% accuracy. On BioDeepAV, it reported 0.99 mAP, 0.99 AUC and 96.77% accuracy. On Vox+FakeAVCeleb, those figures fell to 0.86, 0.88 and 81.59%.
DF-Linear scored 97.87% accuracy on BioDeepAV, above DF-MoE's 96.77%. The two models were tied at 0.99 for mAP and AUC in that comparison.
The supplementary video-only test on Celeb-DF (v2) found an intermediate AUC for DF-MoE: 0.834, while the listed comparator AUCs ranged from 0.789 to 0.875.
Component tests add detail to the headline result
Within MAVOS-DD's open-set full evaluation, the complete cue configuration reported 0.98 mAP, 0.98 AUC and 92.95% accuracy. The listed AVFF+Effort configuration reported 0.96, 0.96 and 88.58%, respectively.
A separate PolyGlotFake comparison examined the CRO component. The configuration with CRO reported 0.929 mAP, 0.941 AUC and 98.49% accuracy, compared with 0.919, 0.936 and 98.42% without CRO. Each metric was higher in the with-CRO configuration, but the differences were small.
The supplement also records both directions of classification error: a real video was labelled fake and a fake video was labelled real. The authors attributed those examples to AV Transformer/AVFF or emotion features.
A sizeable pipeline with a limited trainable core
The reported training setup used 10 epochs. DF-MoE and DF-Linear used AdamW with a learning rate of 10^-4 and batch size 4. DF-MoE used a 128-dimensional shared space, six experts, top-two routing, a CRO loss weight of 1 and an auxiliary-loss weight of 0.01.
The complete pipeline was reported at 508.18 GFLOPs and 424.14 million inference parameters. Training updated 19.9 million parameters, and total inference time was reported as 12.98 seconds with one GPU.
DF-MoE was the top reported method across the four MAVOS-DD scenarios, but it was not highest in every comparison: LipFD had higher AVLips accuracy, DF-Linear had higher BioDeepAV accuracy, and the best listed comparator had a higher Celeb-DF AUC. The reported evidence remains tied to the named datasets and evaluation protocols.
The paper reports a public code release at https://github.com/vladhondru25/DF-MoE. It also reports support from the Ministry of Research, Innovation and Digitization, CCCDI–UEFISCDI, BMFTR and DFG grants or projects.
Paper data and sources
Original title: DF-MoE: Generalizable Deepfake Detection via Multimodal Sparse Mixture-of-Experts
Authors: Vlad Hondru, Florinel Alin Croitoru, Iuliana Georgescu et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-24
DOI: Not available
Original paper · Full text