Research map / Vision-language and multimodal / Cross-modal learning
Investigating Self-Supervised Representations for Audio-Visual Deepfake Detection
Dragos-Alexandru Boldisor, Stefan Smeu, Dan Oneata, Elisabeta Oneata
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use pretrained instead of random-init AV-HuBERT audio and video encoders (ablation) | AV-HuBERT (A+V) synchronization · AV1M | AUC (%): 50.0 -> 87.3 | Improved |
| use pretrained instead of random-init features for next-token prediction (ablation) | AV-HuBERT (A+V) · AV1M | AUC (%): 64.4 -> 84.5 | Improved |
| pair Wav2Vec2 with AV-HuBERT (V) rather than CLIP for synchronization | Wav2Vec2 sync · AV1M | AUC (%): 49.7 -> 86.5 | Improved |
| fuse CLIP with AV-HuBERT (A) for next-token prediction | AV-HuBERT (A) NTP · AV1M | AUC (%): 90.6 -> 86.9 | Hurt |
| use transformer head instead of linear head (ablation) | FSFM features · FAVC (train AV1M) | AUC (%): 40.9 -> 72.3 | Improved |
| use transformer head instead of linear head (ablation) | AV-HuBERT (A) features · AV1M (train FAVC) | AUC (%): 99.0 -> 88.8 | Hurt |
| use max pooling instead of average pooling (ablation) | AV-HuBERT alignment, shift 0 · AV1M | AUC (%): 65.3 -> 76.8 | Improved |
| use min pooling instead of average pooling (ablation) | AV-HuBERT alignment, shift 0 · DFE-2024 | AUC (%): 62.7 -> 66.3 | Improved |
| allow temporal shift of 15 frames instead of 0 (ablation) | AV-HuBERT alignment, average pooling · DFE-2024 | AUC (%): 62.7 -> 75.6 | Improved |
| allow temporal shift of 15 frames instead of 0 (ablation) | AV-HuBERT alignment, min pooling · AV1M | AUC (%): 58.1 -> 56.1 | Hurt |
| use layer 9 instead of the last layer of AV-HuBERT (ablation) | SpeechForensics alignment · AV1M | AUC (%): 67.1 -> 68.6 | No clear effect |
| use BRAVEn (V) visual features instead of Wav2Vec2 audio features | linear probe · Average over FAVC, AVLips, DFE | AUC (%): 71.6 -> 90.5 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- AVFF — AVFF: Audio-Visual Feature Fusion for Video Deepfake Detection (CVPR 2024) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- AVFF: Audio-Visual Feature Fusion for Video Deepfake Detection (CVPR 2024 · similarity 0.92)
- Circumventing Shortcuts in Audio-visual Deepfake Detection Datasets with Unsupervised Learning (CVPR 2025 · similarity 0.91)
- Multi-modal Deepfake Detection via Multi-task Audio-Visual Prompt Learning (AAAI 2025 · similarity 0.90)
- Lips Are Lying: Spotting the Temporal Inconsistency between Audio and Visual in Lip-Syncing DeepFakes (NeurIPS 2024 · similarity 0.88)
- Standing on the Shoulders of Giants: Reprogramming Visual-Language Model for General Deepfake Detection (AAAI 2025 · similarity 0.87)
- I Can Hear You: Selective Robust Training for Deepfake Audio Detection (ICLR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
