Face Sure

A Deepfake Video Analyzer
Detecting deepfake-based cyber threats using deep learning and web deployment

🔍 Overview

Face Sure is a deepfake detection system designed to identify manipulated videos that pose cyber-security threats such as identity impersonation, misinformation, and social engineering attacks.

The project focuses on video-level deepfake detection, combining:

• A deep learning model trained on the FaceForensics++ dataset
• A real-time, web-based interface for video analysis
• A deployment-ready pipeline suitable for practical cyber-defense use cases

🎯 Key Objectives

• Detect deepfake videos instead of isolated images
• Use a modern, efficient deep learning model
• Provide a user-friendly web interface for real-time analysis
• Bridge the gap between academic research and real-world deployment

📂 Dataset

• Dataset: FaceForensics++ (FF++) - C23 compression
• Content:
  • Real videos
  • Manipulated videos:
    • DeepFakes
    • Face2Face
    • FaceSwap
    • FaceShifter
    • NeuralTextures

Dataset Choice Rationale

• Industry-standard benchmark for deepfake research
• Realistic compression similar to social media platforms
• Widely used in foundational deepfake detection papers

⚠️ Note:
Instead of extracting and cropping faces manually, this project uses an already face-cropped version of the FaceForensics++ Dataset, significantly reducing preprocessing complexity while maintaining dataset integrity.

🧠 Model Architecture

Implemented Model

• EfficientNet-B0

Why EfficientNet-B0?

• High accuracy with low computational cost
• More efficient than older architectures (e.g., XceptionNet)
• Suitable for real-time inference and web deployment
• Proven performance in recent deepfake detection research

Classification Setup

The model is trained as a 6-class classifier:

• Deepfakes
• Face2Face
• FaceShifter
• FaceSwap
• NeuralTextures
• Original

A final binary decision (Real vs Fake) is derived at video level using probability aggregation.

⚠️ Note: The training notebook is structured to support multiple architectures, but only EfficientNet-B0 was fully trained, evaluated, and deployed in this implementation.

📊 Evaluation Strategy

Frame-Level Evaluation

• Accuracy
• Precision
• Recall
• Confusion Matrix

Video-Level Decision (Core Contribution)

Since real-world attacks occur via videos, the system performs:

1. Frame sampling from uploaded video
2. Frame-wise model prediction
3. Probability aggregation across frames
4. Final decision using:

🚀 Running the Project Web Application

The Face Sure web application is deployed online and can be accessed directly without any local setup.

1️⃣ Open the Website

Visit the live application using the link above.

You will see:

• Face Sure branding
• Detection controls (sliders)
• Video upload section

No login or installation is required.

2️⃣ Upload Videos

• Upload up to 5 videos at a time
• Supported formats:
  • MP4
  • AVI
  • MOV
  • MKV

You can upload:

• Your own custom videos
• Sample videos from this repo
• Videos from the FaceForensics++ 23 Dataset

Each uploaded video will be displayed on the left side of the interface.

3️⃣ Configure Detection Settings

Before analysis, you can adjust the detection parameters:

• Segment Duration (seconds)
  Controls how long each video segment is analyzed.

• Frames Per Second (FPS)
  Controls how many frames are sampled per second.

• Fake Confidence Threshold
  Determines the sensitivity of fake vs real classification.

These settings allow users to balance speed and accuracy.

4️⃣ Automatic Video Analysis

Once a video is uploaded, the system automatically:

1. Samples frames from the video
2. Detects faces using MTCNN
3. Runs EfficientNet-B0 inference on each face
4. Aggregates predictions across frames
5. Produces a video-level decision

No manual action is required after upload.

5️⃣ View Results

For each uploaded video, the results are shown on the right side:

• Video Title (e.g., Video 1: example.mp4)
• Final Decision
  • REAL ✅
  • FAKE 🚨
• Fake Score
• Runtime
• Class-wise confidence visualization

The Fake Score is calculated as:

Fake Score = 1 − P(Original)

A higher fake score indicates a higher likelihood of manipulation.

6️⃣ Interpret the Output

• REAL
  The video is likely authentic.

• FAKE
  The video shows signs of deepfake manipulation.

The per-class confidence bars provide insight into which manipulation type the model finds most likely.

🛡️ Use Case Perspective

Face Sure is designed for:

• Cyber-security awareness
• Deepfake threat analysis
• Educational and research demonstrations
• Practical deployment scenarios

It emphasizes video-level detection, which better reflects real-world cyber threats compared to frame-only approaches.

⚠️ Notes & Limitations

• Performance depends on video quality and face visibility
• Heavily occluded or low-resolution faces may reduce accuracy
• The current deployment uses EfficientNet-B0 only
• Explainability (Grad-CAM) is planned but not yet integrated

🚦 N.B: The trained model, notebooks, and supplementary documents are not included in this repository. If you want the files, contact me anytime.

✅ Summary

Using Face Sure is simple:

1. Open the live website
2. Upload videos
3. Adjust settings (optional)
4. View real-time deepfake detection results

The system runs fully online and requires no local configuration.