XceptionNet is a Deep Learning Algorithm that Detects Face Swaps in Videos
- A group of researchers from Germany have developed an algorithm that detects forgeries in videos
- The deep learning model was trained on over 1,000 videos which resulted in half a million images
- The results so far have been impressive, even on compressed videos
Swapping a person’s face for another’s has become possible in recent times thanks to deep learning. Unfortunately, this has been misused to harm the reputation of people on the internet. It has even posed a problem for biometric systems! And if this trend continues unchecked, how long before people lose trust in the videos they watch?
To combat this, researchers from the Technical University of Munich have developed a deep learning algorithm that potentially identifies forged videos of face swaps on the internet.
The researchers began by curating a dataset of over 1,000 videos that had face swaps and their original versions. The database they ended up creating contained over half a million images of faces that had been manipulated by software. They are calling this database the FaceForensics dataset.
Once the data had been collected, the team trained it’s deep learning neural network model to understand and grasp the difference between the original video and the manipulated one. This algorithm is being called XceptionNet. Once trained, the algorithm was tested against other pre-existing forgery detection approaches.
The results, as you can see in their research paper (link below), are pretty impressive. Even when the video has been compressed to make the task significantly more challenging, the algorithm has produced results to be hopeful about.
You can read the full research paper here.
Our take on this
Something akin to this algorithm was desperately required to wage the battle against face swaps being used for the wrong reasons. In releasing the research paper to the public, the researchers are hoping others also take up the baton and work on this study to make it more accurate and precise.
But there’s also a caveat with this algorithm – it can also potentially be used to improve the quality of the face swaps which will make it harder to detect the fake. Also, as soon as a forgery detection algorithm is launched, the scammers always try to refine their model to stay a step ahead.
As a data scientist, work like this is exciting since it offers a different way of working with image manipulation problems.
Subscribe to AVBytes here to get regular data science, machine learning and AI updates in your inbox!