VerifAI: trained model weights for multimodal deepfake detection

Trained model weights for VerifAI, a multimodal (audio + visual) deepfake video detection framework. This record contains: Final visual model: Xception trained on FaceForensics++ C23, Celeb-DF v2, DeeperForensics-1.0 and FakeAVCeleb v1.2 Final audio model: ResNet-18 on MFCC features, trained on FakeAVCeleb v1.2 Final late-fusion model: logistic regression, decision threshold 0.25. On the FakeAVCeleb test set: accuracy 0.991, balanced accuracy 0.995, ROC-AUC 0.996 (n = 2,115) Baseline and ablation checkpoints, including the FaceForensics++ Xception baseline used in the cross-dataset (out-of-distribution) evaluation ONNX exports used by the in-browser demo at https://vrifai.com Test metrics and reports for every model, a loading example (load_example.py) and SHA-256 checksums (SHA256SUMS.txt) See README.md inside the record for a description of every file, loading instructions and full results. Source code: https://github.com/giuliolabs/VerifAILive demo: https://vrifai.com No dataset videos are redistributed; the training datasets are available from their original authors under their own terms. The weights are released for non-commercial research and educational use (CC BY-NC 4.0). This is a research prototype: benchmark-trained detectors generalise poorly to in-the-wild video, so its outputs should not be used as sole evidence for forensic, legal or journalistic decisions.