DeepfakeDetectionSystem
A final-year project focused on detecting AI-generated facial content using deep learning, designed as a free and open-source solution to promote transparency, accessibility, and reproducibility.
Accuracy
68%
Architecture
ConvNeXt-Base
Learning
Transfer Learning
Latency
< 300 ms
Detection Systems
Image Deepfake Detection
High-accuracy deepfake detection using ConvNeXt with RetinaFace preprocessing and hybrid spatial learning.
- ■RetinaFace-based face detection
- ■ConvNeXt-Base backbone
- ■ImageNet pretrained
- ■Clean dataset pipeline
- ■FastAPI backend
Haar Cascade + CNN Detection
Lightweight pipeline using Haar Cascade for fast face detection combined with CNN classification for real-time deepfake detection.
- ■OpenCV Haar Cascade detection
- ■Fast CPU inference
- ■Low latency system
- ■No GPU required
- ■Hugging Face deployment
Research to Deployment
End-to-end pipeline covering dataset engineering, dual-domain learning (spatial + frequency), hybrid modeling, and real-time deployment.
Problem Formulation
Formulated deepfake detection as a binary classification problem (REAL vs FAKE), focusing on detecting subtle facial manipulation artifacts and ensuring generalization to unseen forgeries.
Benchmark Dataset (FaceForensics++)
Utilized the FaceForensics++ dataset containing real and manipulated videos. Extracted frames, ensured label correctness, and balanced real vs fake samples for robust training.
Frame Extraction Pipeline
Extracted frames from videos using Haar Cascade, removed redundant frames, and created a structured dataset pipeline suitable for deep learning workflows.
Dual Face Detection Strategy
Implemented Haar Cascade for high-accuracy face localization and Haar Cascade for lightweight, real-time detection, enabling both performance and deployment flexibility.
Face-Centric Preprocessing
Cropped detected faces, resized to 224×224, normalized pixel values, and organized into train/validation/test splits (70/15/15).
Data Augmentation
Applied augmentations such as horizontal flips, rotations, and brightness variations to improve model robustness against real-world variations.
Spatial Feature Learning (ConvNeXt)
Fine-tuned a pretrained ConvNeXt-Base model on ImageNet to learn spatial inconsistencies like texture artifacts, blending errors, and facial distortions.
Frequency Feature Extraction (FFT)
Applied Fast Fourier Transform (FFT) to convert images into frequency domain, capturing hidden periodic patterns and compression artifacts invisible in pixel space.
Hybrid Model Design
Designed a hybrid architecture combining spatial features (ConvNeXt) with frequency-domain features (FFT-based CNN) to improve detection of high-quality deepfakes.
Training Strategy
Trained using AdamW optimizer (LR = 3e-5), partial layer unfreezing, and regularization techniques to stabilize learning and prevent overfitting.
Evaluation & Insights
Observed ~60–70% accuracy with spatial-only model, highlighting limitations. Hybrid approach improves robustness against unseen and high-quality deepfakes.
Deployment (FastAPI + Hugging Face)
Deployed models using FastAPI backend and integrated with a web interface. Lightweight Haar-based pipeline deployed on Hugging Face for real-time inference.
Research & Resources
An overview of the systems, experiments, and research components that form the foundation of the deepfake detection framework.
Image Deepfake Detection (Deployed)
LIVEA ConvNeXt-based spatial deepfake detection model operating on face-cropped images. Designed as a transparent baseline system with documented limitations.
Model Architecture & Training Pipeline
DOCUMENTEDDetailed breakdown of dataset preprocessing, face detection using RetinaFace, model selection, fine-tuning strategy, and evaluation methodology.
FastAPI Inference Backend
OPEN SOURCEProduction-ready inference service exposing the trained model through REST APIs for real-time image analysis and integration.
Frequency-Domain Research (FFT)
IN PROGRESSOngoing research focused on capturing GAN-induced artifacts invisible in RGB space using frequency analysis.
Video Deepfake Detection
LOCKEDTemporal and frequency-aware deepfake detection for videos. Currently under active research and not yet publicly released.
Viva Questions
Key technical questions and detailed explanations related to the deepfake detection system.
These questions are based on real evaluation and viva discussions for deepfake detection systems.
Contact the Research Team
For research discussions, system integration, deployment help, or collaboration opportunities — reach the people behind the model.
We respond to serious research and integration queries only.