Computer VisionStatus: Research
DeepFake Detection
Detecting manipulated faces using transfer learning
PythonTensorFlowMobileNetV2ResNet50KaggleOpenCV
The Problem
Deepfake images are increasingly hard to distinguish from real photos with the naked eye.
The Solution & Architecture
Built a classification notebook using Kaggle's 140k real-vs-fake faces dataset, applying MobileNetV2 and ResNet50 transfer learning to detect manipulated images.
Pipeline Architecture Breakdown
1.Facial Landmark Detection & Alignment (MTCNN / MediaPipe)
2.High-Frequency Artifact Extraction & Texture Analysis
3.ResNet50 & MobileNetV2 Transfer Learning with fine-tuned top dense layers
4.Sigmoid Probability Output with Threshold Optimization for low False Positives
Core Highlights & Functionality
140k Face Dataset Ingestion: Balanced training and validation on diverse facial demographics
Dual Architecture Benchmark: Comparative study between MobileNetV2 (edge efficiency) and ResNet50 (deep feature extraction)
Artifact Analysis: Focus on blending boundaries, eye reflection inconsistencies, and warping artifacts
Reproducible Kaggle Pipeline: Full preprocessing, training callbacks, and confusion matrix evaluations