Deepfakes: Curbing the spread of misinformation

Deepfakes first gained notoriety following a Reddit post outlining how to create them, leading to a glut of falsified videos being created and shared on the platform. Today, access to that particular thread on Reddit has been blocked, but it’s now even easier to access countless apps and tutorials on a range of websites.

Researchers have found that falsified news online is likely to spread faster than accurate information. They speculate this is because there is a bias to sharing negative news over positive news. Falsified news or fake videos are more likely to meet this criteria and therefore are more likely to be shared.

The potential for this technology to spread misinformation across a multitude of areas is vast. Whether through impersonating someone’s voice to commit fraud and fool a colleague, family or friend into transferring money, or damaging the reputation of a business by falsely creating a video from a CEO announcing a major financial loss, or the termination of a partnership with another company.

One option in tackling the deepfake problem involves researching ways to detect deepfakes using ML algorithms. Many deepfakes are created by improving the generation algorithm in sync with the detection algorithm. This can mean that improvements in detection algorithms are quickly followed by improvements in generation algorithms. For example; for a while it was possible to spot a deepfake by tracking the speaker’s eye movements, which tended to be unnatural. Shortly after this method was identified, algorithms were tweaked to include better blinking. It means that these algorithms are creating somewhat of an arms race between creators and detectors. But if researchers can keep finding ‘tells’ in online media, then creators will be forced to continually develop new algorithms.

It’s also possible to add a signature to a video to prove authenticity. Much like watermarks are used on official documents to mark them as legitimate, adding a signature to videos would make it difficult for deepfake systems to tamper with these. Traditionally, incorporating these signatures into live video has been challenging due to variable streaming bit rates, but our innovation teams have recently developed a feature based authentication measure that can be attached to video streams as a signature. It’s informed by visible elements of the footage, before being generated through ML techniques. Viewers or third parties can then authenticate videos by executing the same algorithm and comparing signatures.

Many companies are now providing new and improved methods for detecting deepfakes, but, as is to be expected, the technology used to create them is also evolving. Falsified news isn’t going to disappear overnight, and one of the best things we can do is arm ourselves with the knowledge of what it is and what tools we do have to spot it.

Recently Microsoft announced two different tools for combating deepfakes; a video authenticator (a tool for analysing videos and still photos to generate a manipulation score) and a video authenticity prover (made up of a tool to generate a certificate and a hash for a video, and a tool to check them). But these tools are currently only available to journalists and politicians, and are still in their infancy. A range of parties have been involved in developing these tools - a sign of how difficult it is to develop these things in isolation, even for a huge company like Microsoft.