AI Deepfake & Manipulated Content Verification Tool

Overview

The rapid advancement of generative AI has led to a significant rise in deepfake and manipulated media, including images, videos, and audio. These are increasingly used in fraud, impersonation, misinformation, and social engineering attacks. Elderly and digitally less-aware users are especially vulnerable to such threats.

This project presents a real, end-to-end AI-powered system that verifies the authenticity of images and videos using pretrained deepfake detection models, explainable AI techniques, and asynchronous processing for heavy workloads.

The system is designed as a deployable software product, not a simulation.

---

Key Features

• Image deepfake detection using pretrained models

• Video deepfake detection with temporal aggregation

• Confidence score with human-readable verdicts:

  • Real
  • Suspicious
  • Likely Fake

• Asynchronous video analysis using Redis + Celery

• Explainable AI support for images

• Privacy-first processing (no media permanently stored)

• API-driven architecture ready for web or mobile frontends

---

Supported Media Types

Media Type	Status
Images (JPG, PNG)	Supported
Videos (MP4)	Supported
Audio	Planned
URLs / Links	Planned

---

Model Architecture

Image Detection

• Model: XceptionNet (pretrained on FaceForensics++)
• Input: RGB image
• Output: Manipulation probability
• Reason: Widely used research baseline with strong performance on compressed media

Video Detection

• Frame sampling from video
• Image-level inference on selected frames
• Top-K temporal aggregation to focus on highly manipulated frames
• Final video-level confidence score

This approach follows standard practices used in deepfake detection research and industry systems.

---

Technology Stack

Backend

• Python
• FastAPI
• PyTorch
• OpenCV
• Redis
• Celery

Machine Learning

• XceptionNet (pretrained)
• Torchvision
• TIMM
• NumPy
• Scikit-learn (evaluation and metrics)

Infrastructure

• Asynchronous task queue (Celery)
• Message broker (Redis)
• CPU-based inference (GPU optional)

---

Project Structure

media-detector/
├── backend/
│   ├── api/                # FastAPI routes
│   ├── ml/                 # ML models and evaluation
│   ├── tasks/              # Celery background tasks
│   ├── utils/              # Helpers and utilities
│   ├── main.py             # FastAPI entry point
│   └── celery_app.py       # Celery configuration
├── pretrained/             # Pretrained model weights
│   └── xception_c23.p
└── README.md

---

Setup Instructions

1. Clone the Repository

git clone <repository-url>
cd AI-Deepfake-Manipulated-Content-Verification-Tool/media-detector

---

2. Install Dependencies

pip install -r requirements.txt

Required packages include:

• torch
• torchvision
• timm
• fastapi
• uvicorn
• celery
• redis
• opencv-python
• pillow
• scikit-learn

---

3. Download Pretrained Model Weights

Download the FaceForensics++ pretrained XceptionNet weights:

https://github.com/ondyari/FaceForensics/releases/download/v1.0/xception_c23.p

Place the file here:

media-detector/pretrained/xception_c23.p

---

4. Start Redis

Using Docker:

docker run -d -p 6379:6379 redis

---

5. Start Celery Worker (Windows)

celery -A backend.celery_app.celery_app worker --loglevel=info --pool=solo

---

6. Start FastAPI Server

uvicorn backend.main:app

API documentation will be available at:

http://127.0.0.1:8000/docs

---

Evaluation

The system is evaluated using the FaceForensics++ dataset (C23 compression).

Metrics reported:

• Accuracy
• Precision
• Recall
• F1 Score
• ROC-AUC

Evaluation is performed using real inference, not simulated outputs.

---

Privacy and Security

• Media files are processed in memory or temporary storage only
• No user content is permanently stored
• No personal data collection
• Designed to be compliant with privacy-first principles

---

Limitations

• Audio and live call analysis are not yet implemented
• Video detection relies on image-based models with temporal aggregation
• Performance depends on hardware (CPU vs GPU)