It was a pleasure to participate in the second edition of the Hackathon for Peace, Justice and Security. This journey consisted of 48 hours of fun, networking and coding. Our team won within particular challenge teams and total prize across other challenges.
Hackathon was composed of several challenges that are addressing important social problems. Challenges were provided by international and local companies. Our challenge was provided by UNICRI - United Nations Interregional Crime and Justice Research Institute and offer to develop a solution which is capable to detect manipulated videos,generated by Deepfakes, Face2Face and FaceSwap algorithms. UNICRI published an article about this event. We used Face Forensics Dataset which contains videos that were manipulated. Extracted frames, more specifically extracted face frames, from these videos were used as a primary dataset.
Our solution was based on state-of-the-art neural network architecture EfficientNet for image classification which is smaller and faster than other existing architectures while able to provide better classification accuracy. To simplify user interaction we also developed web application using Dash framework.
You can see one of the fake videos recognized by the trained neural network below:
I would like to thank hackathon organizers for fantastic work and all mentors for their valuable feedback which step-by-step transformed our initial solution to the final shape. We are looking forward to continuing working on this important topic and can wait for the next edition of Hackathon for Peace, Justice and Security.