Hackhers Deepfake Detector

Inspiration

With the rapid rise of AI-generated media, deepfakes have become a serious threat to personal privacy and online trust. We were inspired to build this project after seeing how easily manipulated videos and images can spread misinformation and violate individuals’ privacy. We wanted to create a tool that empowers people to verify media authenticity quickly and easily, giving them more control over their digital presence.

What We Built

We developed a Deepfake Detector that allows users to input a video URL and receive a prediction score indicating the likelihood of the content being manipulated. The app uses a PyTorch-based facial recognition model to analyze frames in the video and identify suspicious patterns typical of deepfakes. The frontend is a simple JavaScript interface where users can enter a video URL and get results instantly.

The workflow is as follows:

User inputs a video URL.
Backend downloads the video and extracts frames.
Each frame is analyzed by the model to detect deepfake signs.
Predictions are aggregated into a single score and returned to the user.

This approach ensures that users can protect their privacy and verify the authenticity of content before sharing or reacting to it.

What We Learned

During this project, we gained experience in:

Integrating PyTorch models with a web frontend.
Handling video data efficiently, including frame extraction and aggregation.
Deploying a full-stack application using Flask and JavaScript.
Navigating challenges in AI deployment, like optimizing prediction speed and handling different video formats.

It also deepened our understanding of computer vision, especially how subtle facial inconsistencies can be detected using neural networks.

Challenges Faced

Video processing: Handling large videos and ensuring fast processing without crashing memory.
Model integration: Combining PyTorch with web APIs while keeping latency low.
Dependency conflicts: Some libraries, like mtcnn and TensorFlow, caused version issues which we resolved by carefully managing Python packages.
Prediction accuracy: Fine-tuning the model to minimize false positives and false negatives required experimenting with frame sampling strategies and aggregation methods.

Despite these challenges, the project successfully demonstrates a working deepfake detection pipeline that enhances digital privacy and raises awareness about the reliability of online media.