DAVID-Net
deepfake detector with separate video and audio analysis
2 min readupdated
Standard deepfake detectors only output a binary "fake or real" verdict and often stumble on dual-modality spoofs where synthetic video is synced with a cloned voice. Developed by researchers led by Moshin Ibna Hossain, DAVID-Net (Decoupled Audio-Visual Forensic Network) takes a more granular approach: it classifies recordings across four quadrants (RVRA, RVFA, FVRA, FVFA) to pinpoint whether the face, the speech, both, or neither has been manipulated. Under the hood, it pairs a VideoMAE ViT-Base encoder for the video stream with WavLM Base+ for audio, separating feature subspaces via orthogonal disentanglement. The tool is built for digital forensics specialists, platform moderators, and media security researchers.
Key features
- 4-quadrant attribution: distinguishes authentic recordings (RVRA), voice spoofs (RVFA), face swaps (FVRA), and dual-modality fakes (FVFA).
- Missing-modality resilience: learnable null tokens let the model handle silent surveillance footage or audio-only call recordings without breaking.
- Tamper boundary localization: a dedicated model head pinpoints the exact frame-level timestamps where manipulation occurs.
- Compact footprint: checkpoint weights take up around 341 MB (75.6M–105.5M parameters depending on the training phase), and the model is released under the MIT license.
What you can do
- Screen suspicious videos: run video interviews and news clips through the CLI script to check whether facial expressions were altered or synthetic speech was dubbed over.
- Analyze phone calls and voice notes: scan standalone audio files (
.wav) for voice cloning even when no video track is available. - Audit silent footage: inspect surveillance camera clips or muted videos for facial manipulation.
Example
from huggingface_hub import hf_hub_download
model_path = hf_hub_download(
repo_id="MIHMahmudEli/david-net-av",
filename="david_net_best.safetensors"
)
config_path = hf_hub_download(
repo_id="MIHMahmudEli/david-net-av",
filename="config.json"
)
Read more
DetailsOn the FakeAVCeleb benchmark, the Phase B version (105.5M parameters) scores 0.978 Clip AUC, 0.982 Video AUC, 0.999 Audio AUC, and 92.4% Macro-F1.
In zero-shot tests without fine-tuning, the model reaches 0.839 on ASVspoof 2019 LA (+19.1% over the baseline) and 0.998 on DeepFakeTIMIT.
Limitation: weak generalization to unseen vocoders in WaveFake (0.113), and output scores are probabilistic (calibrated at T* = 1.18).










Comments