Deepfake Video Detectionv1.0
✦ Project Overview
A real-time deepfake detection system optimized for mobile platforms. It utilizes a lightweight MobileNetV2 architecture with TensorFlow Lite to effectively detect manipulated videos during live calls or media consumption.
✦ Key Features
- ♥Real-time video manipulation detection
- ♥Optimized for mobile devices using MobileNetV2
- ♥Lightweight TensorFlow Lite model integration
- ♥Live video call protection capabilities
✦ Methodology
Focusing on edge-deployment and low-latency inference:
01.
Frame Extraction
OpenCV captures video frames in real-time and applies face detection algorithms to isolate regions of interest (ROI).
02.
Model Optimization
We utilized MobileNetV2, significantly pruned and quantized to run efficiently on mobile CPUs via TensorFlow Lite without sacrificing accuracy.
03.
Inference Pipeline
The model outputs a probability score for 'Real' vs 'Fake' for each frame, and a temporal smoothing filter aggregates scores to prevent flickering predictions.