Source-linked AI summary
DF-MoE: Generalizable Deepfake Detection via Multimodal Sparse Mixture-of-Experts
Vlad Hondru, Florinel Alin Croitoru, Iuliana Georgescu, A. Sophia Koepke, Radu Tudor Ionescu
TL;DR
Audio-visual deepfake detectors need to generalize across changing generation methods, but existing systems often focus on benchmark performance. DF-MoE combines diverse high-level cues from pre-trained audio and visual models with a Mixture-of-Experts backbone and reports superior results across five benchmarks, while also offering feature-attributed decisions.
Problem
Deepfake detection needs detectors that generalize across generation methods because existing benchmark-focused systems may fail on newer manipulations.
Method
DF-MoE extracts diverse high-level audio-visual cues with pre-trained models, integrates them using a Mixture-of-Experts backbone, and uses CRO to organize latent representations for generalization.
Results
DF-MoE outperforms state-of-the-art detectors across five benchmarks in both open-set and cross-domain evaluation scenarios.
Takeaways & Limitations
Combining complementary high-level cues through DF-MoE supports generalizable deepfake detection and feature-attributed interpretations of predictions.
Takeaways & Limitations
Full fine-tuning of the AVFF encoder degrades performance, whereas parameter-efficient adaptation provides superior generalization.
Abstract
from arXiv · showhide
Audio-visual deepfake detection is an actively studied topic, where one of the main challenges is to develop detectors able to generalize across deepfake generation methods. We conjecture that overfitting can be mitigated by extracting multiple high-level cues from the available audio and visual modalities via pre-trained models. We therefore assemble a wide variety of pre-trained models to extract features that encode mouth movements, face parsing, facial expressions, head pose, gaze tracking, heart rate, audio emotion and speech activity. We further integrate both unimodal and multimodal cues via a Mixture-of-Experts (MoE) backbone to detect deepfakes. We perform in-domain and cross-domain experiments on five benchmarks for deepfake detection (MAVOS-DD, AVLips, PolyGlotFake, BioDeepAV, FakeAVCeleb) to compare our framework (DF-MoE) with state-of-the-art methods. Our results indicate that DF-MoE obtains superior deepfake detection results, surpassing all competing methods. We release our code at https://github.com/vladhondru25/DF-MoE.
1 Introduction
DF-MoE addresses the challenge of generalizing audio-visual deepfake detectors across generation methods by combining diverse high-level cues with a sparse Mixture-of-Experts framework. It reports strong open-set and cross-domain performance while supporting more interpretable decisions.
- Motivation: Deepfake detectors must generalize beyond the generation methods represented in existing benchmarks.Reported benchmark accuracy can reach 99%, yet newer and more advanced manipulation methods may remain insufficiently evaluated.
- Approach: DF-MoE extracts multiple high-level audio and video cues through pre-trained models rather than relying primarily on deepfake-specific fine-tuning.The pipeline uses cues from tasks such as facial expression recognition, audio emotion recognition, head pose estimation, and gaze tracking.
- Results: DF-MoE surpasses state-of-the-art detectors across five benchmarks in both open-set and cross-domain evaluations.The benchmarks are MAVOS-DD, AVLips, PolyGlotFake, BioDeepAV, and FakeAVCeleb.
- Approach: DF-MoE integrates complementary cues with a sparse Mixture-of-Experts transformer in an end-to-end detection pipeline.The framework combines high-level semantic features from multiple modalities through specialized experts.
- Approach: CRO is introduced to improve generalization across deepfake generation methods by organizing latent representations around learnable class anchors.The objective contracts samples toward their class anchors and repels the anchors beyond a margin, without expensive hard-sample mining.
- Interpretability: DF-MoE supports interpretable decisions by identifying which high-level features contribute to a classification.Examples include unusual facial expressions, inconsistent speech and mouth movements, and unexpected gaze behavior.
2 Related Work
Related work spans image-level, audio-level, and multimodal deepfake detection, with several studies specifically targeting out-of-domain performance. DF-MoE differs by integrating diverse signals from frozen pre-trained models through a learnable Mixture-of-Experts block.
- Detection paradigms: Deepfake detection methods are categorized by input domain into image-level, audio-level, and multimodal approaches.Early efforts commonly employed convolutional neural networks.
- Multimodal detection: Early multimodal frameworks modeled audio-video synchronization using approaches such as late fusion and two-plus-one fusion.Other studies used pseudo-fake generation to detect local audio-visual inconsistencies.
- Generalization: Prior generalization studies used pseudo-fake data or hybrid supervised and reinforcement learning methods to improve cross-domain detection.One study found that naively training on multiple datasets did not improve joint performance and proposed an alternative strategy.
- DF-MoE: DF-MoE extracts diverse signals from frozen pre-trained models and integrates them with a learnable Mixture-of-Experts block.The approach is presented as a way to reduce overfitting and increase out-of-domain generalization capabilities.
3 Method
DF-MoE detects audio-visual deepfakes by extracting diverse high-level cues with pre-trained models, encoding them through specialized adapters, and integrating them with a sparse MoE transformer. Its training combines binary classification with CRO to organize representations around separated real and fake class anchors.
- 3.1 Extracting Visual and Audio Cues: DF-MoE extracts semantic cues spanning head pose, gaze, face parsing, rPPG, facial expression, speech, audio emotion, lip synchronization, and audio-visual correspondence.The cues are obtained from pre-trained models and capture temporal, physiological, affective, speech, and cross-modal signals.
- 3.1 Extracting Visual and Audio Cues: Face detection and tracking form person-specific face tracks that are processed independently through the feature-extraction pipeline.Faces are detected every fifth frame, then matched across consecutive processed frames with DeepSORT to create long-term tracking paths.
- 3.2 Specialized Adapters: Frozen pre-trained models provide the cues, while trainable adapters encode them into a shared embedding space of size h = 128.The adapters use modality-specific sequence encoders, pooling, and cross-attention to produce compatible representations; the lip-sync adapter processes a custom audio-visual transformer, while AVFF features are linearly projected.
- 3.3 Sparse Mixture-of-Experts: A sparse MoE transformer integrates the adapted tokens, with routing regularization intended to prevent experts from over-specializing in particular feature types.The framework processes the resulting tokens with a trainable sparse MoE transformer, and training applies dropout to expert routing.
- 3.3 Sparse Mixture-of-Experts: The training objective combines binary cross-entropy with CRO, which contracts within-class representations, repels real and fake class anchors, and prevents anchor collapse.CRO uses learnable anchors A and B; its no-collapse term pushes anchors away from the origin, while the full objective supports latent-space separation without hard sample mining.
4 Experiments
DF-MoE is evaluated across five benchmarks and multiple in-domain, open-set, and cross-dataset settings. Results show that diverse pre-trained cues and their MoE integration improve detection performance and support interpretable cue attribution.
- Datasets and evaluation: DF-MoE is tested on AVLips, MAVOS-DD, PolyGlotFake, BioDeepAV, and FakeAVCeleb using in-domain, open-set, and cross-dataset evaluations.The experiments assess generalization across datasets and deepfake generation settings.
- Main results: DF-MoE outperforms previous state-of-the-art methods across MAVOS-DD evaluation setups and achieves the best reported AVLips mAP and AUC.On AVLips, LipFD achieves higher accuracy, but it is specifically designed for lip-sync forgery detection, whereas DF-MoE is more general.
- Ablation studies: The AV Transformer and AVFF provide the strongest individual representations, while HP+Gaze, rPPG, emotion, and face-segmentation cues add consistent complementary gains.Every evaluated high-level cue achieves non-trivial standalone performance, and progressively integrating cues improves the full pipeline.
- Cross-dataset evaluation: DF-MoE achieves the best cross-dataset performance on PolyGlotFake, BioDeepAV, and Vox+FakeAVCeleb in AUC and accuracy.The cross-dataset evaluation includes results reproduced from publicly available code for methods marked with an asterisk.
- Efficiency: Inference time is reported stage by stage for videos on a machine with one AMD Ryzen Threadripper 9960X CPU and two Nvidia RTX 5090 GPUs.The measurements use average video duration and timing estimates across 20 videos.
- Interpretability: Attention visualizations show that DF-MoE assigns different cues high importance for skin-tone variation, eye-region artifacts, altered lip movements, and correctly classified real samples.Audio-visual synchronization cues are important for LivePortrait examples, while head pose and gaze receive high importance for eye-region inconsistencies.
5 Conclusion
The paper addresses out-of-domain generalization in audio-video deepfake detection with DF-MoE, which combines cues from several pre-trained models with trainable mixture-of-experts. Its experiments report state-of-the-art results on three challenging datasets, while acknowledging that additional complementary signals may further improve detection.
- Conclusion: DF-MoE addresses out-of-domain generalization by integrating audio-visual cues from several pre-trained models, including head pose, gaze, face segmentation, rPPG, and emotion.The framework uses frozen pre-trained models together with trainable mixture-of-experts for specialized processing of different input signals.
- Conclusion: Each high-level cue provides useful information for deepfake detection, and DF-MoE achieves state-of-the-art results on three challenging datasets.The conclusion also reports state-of-the-art cross-domain performance from an efficient architecture using frozen pre-trained models.
- Future work: Additional pre-trained models extracting complementary signals could further improve deepfake detection and potentially benefit other applications.This is presented as future work rather than a demonstrated result.
7 Supplementary
The supplementary analyses examine hyperparameters, CRO loss, feature combinations, fine-tuning, missing modalities, attention patterns, and failure cases. Together, they characterize DF-MoE’s robustness, feature behavior, and remaining weaknesses.
- Ablation Studies: d = 0.2 and n = 6 yield optimal performance on MAVOS-DD’s open-set full scenario.The explored configurations also significantly surpass the cited state-of-the-art competitors.
- CRO Loss: CRO loss organizes latent representations to improve generalization across deepfake generation methods.Without CRO, samples from different methods spread across wider regions, and some HifiFace fakes become entangled with real samples.
- Semantic Cue Contributions: Adding complementary modalities can improve feature combinations, with AVFF+Effort+HP+Gaze outperforming AVFF+HP+Gaze on PolyGlotFake.Audio-video emotion features alone with AVFF produce a slight performance drop, whereas adding HP+Gaze to that combination performs better.
- Fine-tuning Feature Extractors: Parameter-efficient adaptation outperforms full fine-tuning of the AVFF encoder, while full fine-tuning can degrade generalization.The supplementary comparison attributes this trade-off to the risk of overfitting dataset-specific forgery artifacts.
- Robustness to Missing Modalities: DF-MoE remains competitive on video-only Celeb-DF (v2) when audio is unavailable.The model safely ignores audio-specific features, supported by its omission of positional embeddings from token representations.
- Qualitative Result Analysis: Average feature importance remains consistent across datasets, supporting reliance on domain-agnostic rather than dataset-specific cues.The attention analysis averages scores over 100 real and 100 fake test videos per dataset.
- Failure Cases: Failure cases arise when AV Transformer, AVFF, or emotion features receive disproportionate influence in incorrect predictions.The authors identify balancing feature contributions and improving sensitivity to frozen extractor precision as directions for improvement.
- Specialized Encoders: The supplementary material also presents architectures for the rPPG, face segmentation, head movement, and emotion encoders.These specialized encoder designs document the components used to extract the pipeline’s high-level cues.