DeepFake Detection System

Designed and contributed to a scalable **Deepfake Detection framework** for identifying manipulated media using **CNN- and Transformer-based architectures**. The system provides an extensible pipeline for **training, evaluation, benchmarking, and inference** across widely used deepfake datasets, supporting both research and real-world security applications.

The framework enables standardized benchmarking of multiple state-of-the-art models on **FaceForensics++ (RAW, C23, C40)** and **Celeb-DF**, allowing fair comparison across compression levels and manipulation types.

Key Features:

• 🕵️ End-to-End Detection Pipeline: Video → frame extraction → preprocessing → model inference → classification.
• 🧠 Multi-Model Architecture Support: ResNet, Xception, EfficientNet, MesoNet, GramNet, F3Net, ViT, and M2TR.
• 🎯 High Detection Performance: Achieved up to 95.9% accuracy on FF-DF and 94.7% on Celeb-DF across baseline models.
• 🧬 Frequency & Attention-Based Learning: Captures spatial, temporal, and frequency-domain forgery cues.
• ⚙️ Research-Ready Design: Modular, configurable framework for training, evaluation, visualization, and inference.

Architecture Overview:

• Video Processing Pipeline: Real/Fake video ingestion with frame extraction and preprocessing.
• CNN-Based Models: Learn local spatial artifacts and texture inconsistencies in manipulated frames.
• Transformer-Based Models: Capture global context and long-range manipulation patterns.
• Unified Evaluation Framework: Standardized benchmarking across FF-DF (RAW, C23, C40) and Celeb-DF datasets.

Technical Highlights:

• Implemented in PyTorch with YAML-based experiment configuration
• Advanced preprocessing including face alignment, normalization, and frame sampling
• Supports training, evaluation, visualization, and single-image/video inference
• Integrated performance metrics: Accuracy, AUC, Precision, Recall, and F1-score