Facebook develops way to reverse-engineer and trace deepfakes

Deepfakes are used to spread misinformation on - and tech giants and fighting back

Facebook, in collaboration with the researchers from Michigan State University (MSU), has developed a new technique it claims can detect deepfake images. On top of detection, the researchers and Facebook say the new approach can also reverse-engineer deepfakes to reveal identifying features of the machine learning model used to create them.

The technique could help to trace the creators of fake images and videos in real-world settings, where the deepfake image itself is often the only information available for detectors to work with.

Deepfakes refer to photos, videos or audio files that are manipulated in such a way that they exhibit someone doing/saying something that the person has actually never done or said.

The advancement of AI technology has made it easier than ever to create deepfakes. AI tools can manipulate individual faces based on previous pictures, to create realistic new images. Fakes like these can be used to spread misinformation online.

Many deepfake videos mimicking celebrities and politicians have emerged in recent years.

In 2019, an Instagram user published a deepfake video of Facebook founder Mark Zuckerberg, drawing attention to Facebook's then-laissez-faire deepfakes policy. That same year Deeptrace, an Amsterdam-based security firm, found nearly 14,700 deepfake videos on the internet in June and July, up 84 per cent from under 8,000 in December 2018.

Politicians are a frequent target, especially those of an American and left-leaning persuasion. Barack Obama, Joe Biden and Nancy Pelosi have all been the subject of deepfake videos - sometimes humorously, but often not.

Facebook says that the current technology around deepfakes 'focus on telling whether an image is real or a deepfake (detection), or identifying whether an image was generated by a model seen during training or not (image attribution via 'close-set' classification).

'But solving the problem of proliferating deepfakes requires taking the discussion one step further, and working to understand how to extend image attribution beyond the limited set of models present in training.

'It's important to go beyond close-set image attribution because a deepfake can be created using a generative model that is not seen in training.'

Facebook says its new software can identify the AI model that was used to generate the deepfake in the first place, no matter how innovative the technique is.

The new reverse-engineering technology begins by running a deepfake through a fingerprint estimation network (FEN) to pull out details about the 'fingerprint' - the unique patterns, like minor oddities in the colour spectrum or slight speckles of noise, left by the model on the deepfake.

The researchers estimated fingerprints using different constraints based on properties of deepfake fingerprints found in the wild. They used those constraints to generate a dataset of fingerprints that was utilised to train a model to detect fingerprints it hadn't seen before.

MSU researchers tested the new reverse engineering method on a database of 100,000 deepfake images generated by 100 publicly available generative models. Facebook said the results demonstrated the effectiveness of fingerprint estimation and hierarchical learning.

MSU is expected to open-source the data set, code and trained models in the coming months to help other researchers study the detection and origin of deepfakes.

The new technology from Facebook and MSU comes less than a year after software giant Microsoft released a set of new tools to help combat deepfakes ahead of the 2020 Presidential election.

The company said in September that its 'Video Authenticator' tool can analyse a still image or a video clip to determine whether the media has been edited using AI technology, while a second tool allows video creators to certify that their content was authentic.