Improving Deepfake Detection Algorithms and Solving the Generalization Gap
Deepfake technology is a relatively novel technique for creating or manipulating images or videos. The rapid evolution of deep learning techniques has resulted in a wide range of possibilities for creating such material. The negative effect of this technology is that it is now easier than ever to disseminate disinformation, spread revenge porn, commit financial frauds, hoaxes, or disrupt government functioning. Therefore it is of paramount importance to be able to manipulate Deepfake detection in footage.
Deepfakes are a well-known revolution in the AI applications field. Many software applications use it for fun and/or social media engagement. But this technology has an obvious and potentially enormous dark side. High-profile individuals like politicians and celebrities are at risk of being Deepfaked. With the widespread use of Deepfake content, problems such as manipulation of public opinion, revenge porn, attacks on personal rights, violations of rights of intellectual property, and personal data protection are becoming more common.
The need to distinguish between authentic and Deepfake material has resulted in vast amounts of research on the topic of Deepfake detection. Most Deepfake detection techniques approach Deepfake detection as a binary classification problem. Although this method yields impressive results on benchmark datasets, it fails to generalize to other, out-of-distribution, Deepfake examples. In this project, the aim is to solve this “generalization gap“.
The Project Goals
The goal of this project is to come up with and implement Deepfake detection techniques with excellent generalization performance. There are two types of Deepfakes, face generation techniques (example: styleGAN) or face manipulation techniques (example: face-swapping algorithms or lip-sync algorithms). We aim to detect both types of Deepfakes as being a Deepfake.
The envisioned solution could be either one of the following:
A (set of) models with excellent generalization performance
A (set of) models that can classify into an “unknown” class
Any other approach is deemed appropriate in solving the generalization gap.
The envisioned solution should work on images and individual video frames. We only consider faces.
DuckDuckGoose has a large dataset consisting of both open source and proprietary Deepfake detection datasets. These datasets contain examples (either images or video) of real people and their Deepfaked counterparts. These datasets are available for this project. Different open source Deepfake datasets also exist.
Why join? The uniqueness of Omdena AI Challenges
A collaborative experience you never had in your working life! For the next eight weeks, you will not only build AI solutions to make a real-world impact but also go through an entire data science project lifecycle. This covers problem scoping, data collection, and preparation, as well as modeling for deployment.
And the best part is that you will join a global and collaborative team of changemakers. Omdena AI Challenges are not a competition or hackathon but a real-world project that will take your experience of what is possible through collaboration to a new level.