eftt/VideoMae-ffc23-deepfake-detector

🧾 Model Card — VideoMAE-DeepFake-Detector-v1 🧠 Model Overview

VideoMAE-DeepFake-Detector-v1 is a fine-tuned video deepfake detection model trained to distinguish between authentic and manipulated facial videos. The model builds upon the pretrained VideoMAE architecture and adapts it for binary classification of real versus synthetic videos.

The base model was originally trained on large-scale video action datasets, enabling strong spatiotemporal feature understanding. It was further fine-tuned on the FaceForensics++ dataset to detect visual artifacts, temporal inconsistencies, and manipulation signatures commonly found in deepfake videos.

By leveraging transformer-based video representation learning, the model captures both frame-level visual cues and motion patterns across time, allowing it to identify subtle manipulations that traditional image-based detectors may miss.

The model is designed for applications in media verification, misinformation detection, and AI-generated content monitoring.

🏗️ Training Details

Base Model: MCG-NJU/videomae-base-finetuned-kinetics

Framework: Hugging Face Transformers + PyTorch

Training Hardware: NVIDIA T4 GPU (Kaggle)

Epochs: 15

Batch Size: 4

Learning Rate: 2e-5

Optimizer: AdamW

Video Sampling: 16 frames per video clip

Resolution: 224 × 224

Training Strategy:

Transfer learning with partial freezing:

~70% of VideoMAE backbone layers frozen

Final transformer layers + classifier head fine-tuned

Dataset: FaceForensics++ (C23 compression level)

Classes:

🟢 Real Video 🔴 Deepfake Video

📊 Dataset Description

The model was trained using the FaceForensics++ dataset, a widely used benchmark for deepfake detection research.

FaceForensics++ contains manipulated videos generated using multiple facial manipulation techniques, including deepfake generation and facial reenactment.

For this model version, training used a subset consisting of:

Original videos (real)

Deepfakes manipulation videos (fake)

Each video was processed by sampling 16 frames uniformly across its duration to capture both spatial and temporal artifacts.

Label Description Real Authentic unmodified video Fake Video manipulated using deepfake synthesis techniques 🎯 Evaluation Metrics

Evaluation was performed on a held-out validation split of the dataset.

Metric Score Train Loss 0.303 Validation Loss 0.506 Accuracy 88.0% F1 Score 0.742 AUC 0.836

✅ The model demonstrates strong ability to distinguish between authentic and manipulated videos using temporal visual patterns.

💬 Example Usage import torch import numpy as np from decord import VideoReader, cpu from PIL import Image from transformers import VideoMAEForVideoClassification, VideoMAEImageProcessor

model = VideoMAEForVideoClassification.frompretrained( "yourusername/videomae-deepfake-detector" )

processor = VideoMAEImageProcessor.frompretrained( "yourusername/videomae-deepfake-detector" )

def loadvideoframes(videopath, numframes=16): vr = VideoReader(videopath, ctx=cpu(0)) totalframes = len(vr)

indices = np.linspace(0, totalframes - 1, numframes).astype(int) frames = vr.get_batch(indices).asnumpy()

return [Image.fromarray(f) for f in frames]

@torch.nograd() def predict(videopath): frames = loadvideoframes(videopath) inputs = processor(frames, returntensors="pt")

outputs = model(**inputs) probs = torch.softmax(outputs.logits, dim=1)[0]

return { "real": float(probs[0]), "fake": float(probs[1]) }

print(predict("sample_video.mp4"))

Output example:

{'real': 0.96, 'fake': 0.04} 🧩 Intended Use

Deepfake detection in video content Media authenticity verification AI-generated video detection pipelines Research on manipulated media detection Integration into misinformation monitoring systems

⚠️ Limitations

The model was trained on a subset of FaceForensics++ and may not generalize perfectly to unseen deepfake generation techniques.

VideoMAE — Base architecture for video representation learning FaceForensics++ — Deepfake detection dataset benchmark Hugging Face Transformers — Training and deployment framework