Could AI help spot a fake Donald Trump?

One of the major use cases for AI is object recognition for machine vision and other applications. The number of uses to which object recognition can be put is huge: think of how much of human life is based on the ability to identify what we see – both ‘in the flesh’, and in reproductions such as photographs, paintings, drawings and videos.

Some images are designed to be symbolic: signage uses representations that abstract the essential attributes of a thing in a way that can be easily interpreted by a human. Sometimes the abstraction and final representations draw on more than data that is just pictorial, but could be cultural too. For instance, identifying ladies’ and gents’ toilets from their signs assumes the viewer understands the cultural significance of the A-line skirt. At an even higher level of abstraction, a “no entry” sign is clear, once learnt, but without being told what it is, perhaps not so easy to interpret. Bar codes are pretty much impossible for a human to interpret, yet extremely simple for automated systems.

Different machine vision tasks require different approaches, and different combinations of technologies. No-one would suggest that AI is needed for supermarket barcode scanning – though the ability to reliably identify the variety of an unlabelled apple might be very useful. In fact, the identification of foodstuffs is an active area of applied machine vision research and development. In food processing, the control of ingredients through the process is more readily done using machine vision than other process control methods as it’s not easy to attach markers to foodstuffs to track them through a safe, hygiene-first, controlled process. In this application, object recognition, using machine learning, becomes simply another sensor, and the systems that are built for it – such as those from Irida Labs, JM Vistec and Industrial Vision Systems – must be optimized for that specific use so they are cost-effective.

Machine vision for autonomous vehicles is similar to the process control example, but with even greater demands on accuracy (Is that a pedestrian or a billboard? Is it moving into the road or away from the road?) and many more objects to identify and classify, and more complex integration into other systems.

Fake news?

But consider another role for AI in object recognition: spotting a photograph or video where part of the image has been manipulated. Such a challenge faces news organisations on a regular basis: sensitivity over “fake news” means responsible publishers are on heightened alert to potential manipulation. This too requires certainty in the decision, just as the food processing system must ensure it prevents strawberries being added to apple yoghurt. Can AI help? Yes, but what’s needed from the AI is very, very much more complex. A human process is probably something like that shown in the figure, which also shows where a machine-learning system might be trained to identify images of specific, high-profile individuals. But trained to spot a fake? That’s different.

Spotting a fake image of Donald TrumpTo identify fake images, we – or an AI – are not looking at the image in the same way as we do when we are identifying what the image is or represents. Rather, we are looking for a set of artefacts (missing shadows, extra limbs, distorted lines that should be straight, etc) and other picture- and non-picture-related details (excessive file size, image compression, absence of the expected optical aberrations, etc) that humans cannot always spot, and which require machine forensic analysis. And there’s something more: was the person in the picture actually in that place at that time? Could he or she physically or – big call this – emotionally do that? I reckon the AI needed to reliably detect fakes is going to need to be quite a lot more advanced than the industrial machine vision systems currently being used.

Add this: