Captchas – those groups of distorted letters and numerals, set against a grainy background, that we are frequently challenged by when we sign up to a website.
The website is simply trying to protect itself against robot software that mimics people and signs up for an account, usually in order to do something that will harm the website, directly or indirectly. The robots are not on the website’s side.
Of course the idea is that humans can recognise the distorted characters but robots can’t, so the robots fail the challenge.
In practice this leads to an arms race. People writing the robots get cleverer and the people designing the Captchas respond by making the Captchas harder to recognise, and so it goes.
In the middle of this war between goodies and baddies the person with low vision becomes collateral damage. Even people with normal vision begin to find it hard to pass the Captcha test. As a result Captchas are widely detested.
Recently there have been lots of media reports about what looks like a breakthrough in artificial intelligence. The start-up Vicarious claims their software can recognise Captchas with 90 percent accuracy. They’ve tested it with Google and Yahoo and PayPal Captchas.
The technology used is one that attempts to imitate the way the human brain works in visual recognition. It has layers of artificial brain cells – the neurons. The layer at the front receives the input. The neurons there may ‘fire’ and send signals to neurons they are connected to in the second layer. The strength of the stimulus those neurons receive depends on the strength of the connections, which can change over time.
If the stimulus a neuron in the second layer receives is sufficient it will fire in its turn and propagate a signal to further neurons in the following layer. From that layer the wave of activation will move forward to the next and so on. The effect of the original stimulus will ripple through the layers of the network and eventually reach an output layer where hopefully something will have been recognised.
The idea that the pioneers studying these ‘neural nets’ had was that a net could learn. You would train it with examples and each example would modify the strengths of some of the connections from one neuron to another, so next time the signals would propagate through the network a bit differently. With enough training it would get as good as a real brain, it was argued.
It’s a lot harder than that.
It turns out that if you have enough layers and enough artificial neurons (which are actually simulated in software, they aren’t literally tiny blobs of jelly), and apply enough training you can create systems that are useful in limited applications.
But we are very far indeed from producing any system that can cope with general recognition tasks as well as the brain-eye systems found in real-world animals.
I think it’s possible in principle to produce artificial vision recognition systems that are as good as natural ones, and one far-off day perhaps we will, but we are a long way from it at present.
Back to the Captcha cracker. Has it really rendered the Captcha
obsolete and will we see them disappear in short order, to be replaced by something completely different?
I doubt it very much indeed. I think the claim that Captchas
have been cracked is publicity for the company making it, but in practice quite a small modification to the Captcha will defeat the recognition system. That’s my view anyway.
Over the years there have been many extravagant claims for neural networks but the practical success has fallen well short of the hype. So I think it’s too soon to rejoice in the overthrow of the Captcha.
For all that this latest this latest development represents a real step forward in the arms race and Captcha recognition software will continue to improve. In response Captcha will get more difficult and all this will make it harder for real people!
When the robots get better than all of us humans, then the Captcha really will be dead and websites will need to use a completely different technology to distinguish real people from robots.
What then? Who knows, it could turn out worse!