B Mahesh Raj
Coimbatore, Tamil Nadu
Ettimadai, Tamil Nadu
Developing a robust Deepfake Detection Web Dashboard, leveraging parallel processing and ensemble models. User-friendly interface, multiple detectors, and a human-in-loop system ensure efficient analysis, prevention of conflicting outputs, and continuous improvement in the fight against deepfake con ...learn more
Project status: Under Development
oneAPI, Artificial Intelligence
Intel Technologies
oneAPI,
Intel Python,
Intel powered laptop
The Deepfake Detection Web Dashboard is an innovative project designed to combat the escalating threat of deepfake content. Utilizing a curated list of detectors, parallel processing, and ensemble models, it offers a robust solution for efficient and accurate detection. The user-friendly interface ensures accessibility, while the human-in-loop system allows continuous improvement. This tool plays a crucial role in addressing misinformation, preserving the integrity of information, and fortifying democratic processes. By providing a comprehensive and adaptive platform, it contributes significantly to the ongoing battle against the proliferation of deceptive media in the digital landscape.
We introduce a Deepfake Detection Web Dashboard that enables users to upload videos for scrutiny to determine if they are deepfakes. The submitted videos undergo analysis by a carefully selected set of tested and verified deepfake detectors. Each detector utilizes distinct techniques, possessing its own set of strengths and weaknesses. The individual results from these detectors are intelligently aggregated by an ensemble model, calibrated based on the historical performance of the employed detectors.
The system employs a list of tested, and verified deepfake detectors with each utilizing distinct techniques, having its strengths and weaknesses. An ensemble model is developed and trained to intelligently aggregate probabilities from individual detectors to produce a single output probability. The aggregation model helps prevent conflicting outputs and provides a comprehensive overview, considering the unique strengths and weaknesses of each detector. The final results are presented to the user through graphical representation. Combining results from multiple detectors makes adversarial attacks more difficult, as they typically exploit the weaknesses of individual models.
Each detector in the list has two methods: Video Preprocessing Method: Converts uploaded videos into tensors. Inference Method: Applies deepfake detection algorithms to the preprocessed tensors, yielding probabilities for each video. Detectors are designed with a split into 2 approach, allowing efficient code reuse.
To address computational overhead, the system utilizes parallel processing (Intel oneAPI –Thread building block) for efficient video processing. Intel AI analytic tools and optimized libraries enhance the overall efficiency and reduce processing time. Mediapipe library is employed for face detection and cropping during the preprocessing phase, accelerating processing speed.
Multi-face Handling: For videos with multiple faces, various faces are presented to the user for selection, ensuring compatibility with algorithms trained for individual faces.
Human-in-Loop System: Enables expert analysis of stored video inputs and outputs for ongoing model performance improvement also facilitates continuous analysis and enhancement based on real-world data.
Parallel Processing: - Intel oneAPI – Thread Building Block - Optimized Libraries: - Intel optimized versions for PyTorch, TensorFlow, scikit-learn - Intel extension for scikit-learn The proposed solution integrates cutting-edge technologies and methodologies to create a robust and adaptable deepfake detection model. Through parallel processing, code reusability, and ensemble models, the web dashboard aims to offer a comprehensive tool for users, addressing challenges in the dynamic landscape of deepfake detection.