Can facial recognition detect children's parents?
Can facial recognition show similarities between parents and their children?
Listener Paul got in touch via the webform to ask this question, which Tim Revell answered...
Tim - Well, the answer is yes and no. So face recognition systems, the way that they work the most common types use a technology called convolutional neural networks. And these are a type of algorithm that split up an image into lots and lots of different component parts. And then it analyses those different components. So for example, one part of the neural network might focus on your eye, it might then look at the shape, it might look at the color and that could easily then be compared to another person. So with that technology we have at the moment, we could easily make a comparison between two images of say you and one of your parents. But I think what's interesting about this is what the algorithm will pick up, it could be completely different to the similarities that a human might pick up. So I have a printed off at home prop here that I'm going to show Ella. Ella, who is in these two pictures?
Ella - Reese Witherspoon?
Tim - Reese Witherspoon. Correct. And could you describe the difference between the one on the left and the one on the right.
Ella - One appears to have some glasses on and the other does not.
Chris - Okay. So yes, you've got two pictures side by side and one of them has got, it's clearly pasted on graphic art. You've stuck on a pair of fairly cheap looking glasses frames, not designer. She would definitely not wear anything like that.
Tim - Yeah, exactly. But you can see that this is the same image and then there's some fairly dodgy glosses that have been photo-shopped on over the top. But if you ask a neural network what these two images are, the one on the left, it says Reese Witherspoon, the one on the one on the right. It says it's Russell Crowe.
Chris - How? Where's the beard?
Tim - Well, exactly. So to a human eye there is no way you would ever say this is Russell Crowe.
Chris - Really the best devices say that's Russell Crowe?
Tim - The best devices say this is Russell Crowe, and it's because of these glasses. The way these glasses have been chosen, the little sort of pixel pattern on there, you can see it looks a little bit like a dodgy tortoise shell glass, but the way an algorithm analyses an image. It looks at these individual pixels and it picks up on things that are not like how humans see things at all. What that means is that though you could get face recognition to show similarities between a person and their relatives, it may pick up on things that just don't seem to resemble the similarities that we as humans see in other people.
Chris - If we know that it's being fooled in that way, can't we just write better programs?
Tim - Thing is, it's very difficult to do that. This is called an adversarial attack where you tweak the image so that it confuses a neural network, and you can improve systems against these types of attacks, but it's really, really difficult. And every time someone tweaks their algorithms so that it's good at defending against one sort of attack, someone comes up with a slightly different one. And actually this is quite worrying because if you're a driverless car for example, and someone puts one of these sorts of stickers on a road sign, suddenly you might think a stop sign is actually a go sign, and that could be a big, big issue going forwards.
Chris - So how do they plan to surmount this?
Tim - They don't really have a good answer for this yet. Part of it is that their hope that you get many different opportunities to look at a stop sign or a go sign when you're driving along. So it's not just one image like you get here of Reese Witherspoon. Another way that you might do it is you have an extra part of the system. So, for example, stop signs could have something embedded in them that the car can bounce off and it gets some electronic signal that says, Oh, this is a red light at the moment. But it is something that is currently still up for grabs. You know, they need to work out how to solve it.
Chris - I'm really quite shocked Tim, it's not often I'm shocked by something like that. I didn't realise it was so vulnerable.
Tim - Yeah. Well it's one of those things that you know, that the technology is moving so fast. It's only a few years ago that even recognising Reese Witherspoon was too difficult for an algorithm. So it's come a long, long way. But, for it to be truly safe on the roads, it's going to have to go a little bit further still.