DeepSafe
- Role
- Full-Stack / AI Project
- Timeline
- Final Year Project
- Status
- Completed
- Category
- AI / Machine Learning
A full-stack deepfake detection application that analyzes both video and audio content using dedicated deep-learning pipelines.
DeepSafe is a full-stack web application that detects deepfake manipulation in both video and audio files. It runs two separate deep-learning models - XceptionNet for video and AASIST for audio - through a Django REST backend, with a React/TypeScript frontend that displays frame-level confidence scores, interactive charts, and exportable PDF reports.
The problem
Deepfake content - synthetically generated video faces and cloned voices - has become increasingly difficult to detect by eye. Most detection tools target only one media type, require specialist environments to run, or produce results that are hard to interpret. DeepSafe brings both detection pipelines into a single accessible web interface with a clear, explainable output.
Approach
Video detection
Uploaded videos are decoded frame-by-frame using OpenCV. Each frame is passed through XceptionNet - a CNN trained on the FaceForensics++ dataset - which outputs a per-frame confidence score. The scores are aggregated and visualised as a bar chart and timeline in the dashboard. Repository-reported accuracy: approximately 95%.
Audio detection
Audio files are analysed using AASIST (Audio Anti-Spoofing using Integrated Spectro-Temporal Graph Attention Networks), a PyTorch model trained on the ASVspoof 2019 LA dataset. It detects voice cloning and text-to-speech synthesis. Repository-reported accuracy: 99.17%. These figures reflect model evaluation on the stated datasets and should be understood as research benchmarks rather than guarantees in all real-world conditions.
API and authentication
A Django REST API handles file upload, queues detection jobs, returns structured JSON results, and manages user sessions. Users can view their full detection history and download individual analyses as PDF reports or export summary data as CSV.
Frontend
The React 18 + TypeScript interface is built with Vite and Tailwind CSS. Recharts renders confidence-score visualisations. shadcn/ui components (built on Radix) provide the accessible component base. jsPDF generates downloadable report files client-side.
Infrastructure
The application is containerised with Docker for consistent local and production deployment. PostgreSQL is the primary database. Static files are served via WhiteNoise; production WSGI is Gunicorn.
Stack
Backend
AI Models
Frontend
What it changes
- Video deepfake detection via XceptionNet: ~95% accuracy on FaceForensics++ (model evaluation benchmark).
- Audio spoof detection via AASIST: 99.17% accuracy on ASVspoof 2019 LA (model evaluation benchmark).
- Frame-by-frame video analysis with per-frame confidence scores visualised as charts.
- Authenticated user dashboard with detection history, CSV export, and PDF report generation.
- Admin dashboard for system statistics and user management.
- Fully containerised deployment with Docker.
Model accuracy figures are as reported in the repository and reflect evaluation on the FaceForensics++ and ASVspoof 2019 datasets. DeepSafe is an academic Final Year Project and is not intended as a certified forensic tool.