metadata
title: DEEP FAKE DETECTOR
emoji: 🎭
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: 6.14.0
app_file: app.py
pinned: true
license: mit
short_description: Multimodal deepfake detection across 5 modalities
Multimodal Deepfake Detection System
A modular AI system for detecting fake content across 5 modalities — text, audio, image, video, and lip-sync — using hybrid CNN-Transformer-LLM architectures.
Live Demo
Try the live system using the tabs above:
- Text Verification — Detects fake news using SERP API + DeBERTa + Llama-3.3
- Audio Forensics — Identifies AI-cloned voices using Wav2Vec2 + AST + CLAP
- Image Analysis — Detects deepfake images using ResNet-18 + Vision Transformer
- Video Forensics — Analyzes video deepfakes with multi-frame voting + OpenCLIP
- Multi-Modal Lip-Sync — Detects audio-video mismatch using LSE-Net + Claude reasoning
Performance
| Modality | Accuracy |
|---|---|
| Text | 99% |
| Image | 92% |
| Audio | 91% |
| Video | 91% |
| Lip-Sync | 89% |
Tech Stack
- Frameworks: PyTorch, Hugging Face Transformers, Gradio
- Vision Models: Vision Transformer (ViT), ResNet-18, OpenCLIP
- Language Models: DeBERTa-v3, Llama-3.3, Claude 3.5 Sonnet
- Audio Models: Wav2Vec2, AST, CLAP, Whisper
- Datasets: FaceForensics++, FakeAVCeleb, ASVspoof
About
This system was developed as part of MS thesis research at GIFT University, Pakistan.
Author: Aneela Pervez
Institution: GIFT University, Department of Computer Science
Supervisor: Dr. Muhammad Ziad Nayyer
Links
- GitHub Repository: https://github.com/ANEELA-PERVEZ
- LinkedIn: https://linkedin.com/in/aneela-pervez-95472b310
- Paper: Coming soon on arXiv
Citation
If you use this system in your research, please cite: