As AI increases in corporate, social, economic and political importance, it is becoming a big target for activists and I think there are too many vulnerabilities. I think we should be seeing a lot more articles than we are about what developers are doing to guard against deliberate misdirection or corruption, and already far too much enthusiasm for make AI open source and thereby giving mischief-makers the means to identify weaknesses.
I’ve written hundreds of times about AI and believe it will be a benefit to humanity if we develop it carefully. Current AI systems are not vulnerable to the terminator scenario, so we don’t have to worry about that happening yet. AI can’t yet go rogue and decide to wipe out humans by itself, though future AI could so we’ll soon need to take care with every step.
AI can be used in multiple ways by humans to attack systems.
First and most obvious, it can be used to enhance malware such as trojans or viruses, or to optimize denial of service attacks. AI enhanced security systems already battle against adaptive malware and AI can probe systems in complex ways to find vulnerabilities that would take longer to discover via manual inspection. As well as AI attacking operating systems, it can also attack AI by providing inputs that bias its learning and decision-making, giving AI ‘fake news’ to use current terminology. We don’t know the full extent of secret military AI.
Computer malware will grow in scope to address AI systems to undermine corporate value or political campaigns.
A new route to attacking corporate AI, and hence the value in that company that relates in some way to it is already starting to appear though. As companies such as Google try out AI-driven cars or others try out pavement/sidewalk delivery drones, so mischievous people are already developing devious ways to misdirect or confuse them. Kids will soon have such activity as hobbies. Deliberate deception of AI is much easier when people know how they work, and although it’s nice for AI companies to put their AI stuff out there into the open source markets for others to use to build theirs, that does rather steer future systems towards a mono-culture of vulnerability types. A trick that works against one future AI in one industry might well be adaptable to another use in another industry with a little devious imagination. Let’s take an example.
If someone builds a robot to deliberately step in front of a self-driving car every time it starts moving again, that might bring traffic to a halt, but police could quickly confiscate the robot, and they are expensive, a strong deterrent even if the pranksters are hiding and can’t be found. Cardboard cutouts might be cheaper though, even ones with hinged arms to look a little more lifelike. A social media orchestrated campaign against a company using such cars might involve thousands of people across a country or city deliberately waiting until the worst time to step out into a road when one of their vehicles comes along, thereby creating a sort of denial of service attack with that company seen as the cause of massive inconvenience for everyone. Corporate value would obviously suffer, and it might not always be very easy to circumvent such campaigns.
Similarly, the wheeled delivery drones we’ve been told to expect delivering packages any time soon will also have cameras to allow them to avoid bumping into objects or little old ladies or other people, or cats or dogs or cardboard cutouts or carefully crafted miniature tank traps or diversions or small roadblocks that people and pets can easily step over but drones can’t, that the local kids have built from a few twigs or cardboard from a design that has become viral that day. A few campaigns like that with the cold pizzas or missing packages that result could severely damage corporate value.
AI behind websites might also be similarly defeated. An early experiment in making a Twitter chat-bot that learns how to tweet by itself was quickly encouraged by mischief-makers to start tweeting offensively. If people have some idea how an AI is making its decisions, they will attempt to corrupt or distort it to their own ends. If it is heavily reliant on open source AI, then many of its decision processes will be known well enough for activists to develop appropriate corruption tactics. It’s not to early to predict that the proposed AI-based attempts by Facebook and Twitter to identify and defeat ‘fake news’ will fall right into the hands of people already working out how to use them to smear opposition campaigns with such labels.
It will be a sort of arms race of course, but I don’t think we’re seeing enough about this in the media. There is a great deal of hype about the various AI capabilities, a lot of doom-mongering about job cuts (and a lot of reasonable warnings about job cuts too) but very little about the fight back against AI systems by attacking them on their own ground using their own weaknesses.
That looks to me awfully like there isn’t enough awareness of how easily they can be defeated by deliberate mischief or activism, and I expect to see some red faces and corporate account damage as a result.
This article appeared yesterday that also talks about the bias I mentioned: https://techcrunch.com/2016/12/10/5-unexpected-sources-of-bias-in-artificial-intelligence/
Since I wrote this blog, I was asked via Linked-In to clarify why I said that Open Source AI systems would have more security risk. Here is my response:
I wasn’t intending to heap fuel on a dying debate (though since current debate looks the same as in early 1990s it is dying slowly). I like and use open source too. I should have explained my reasoning better to facilitate open source checking: In regular (algorithmic) code, programming error rate should be similar so increasing the number of people checking should cancel out the risk from more contributors so there should be no a priori difference between open and closed. However:
In deep learning, obscurity reappears via neural net weightings being less intuitive to humans. That provides a tempting hiding place.
AI foundations are vulnerable to group-think, where team members share similar world models. These prejudices will affect the nature of OS and CS code and result in AI with inherent and subtle judgment biases which will be less easy to spot than bugs and be more visible to people with alternative world models. Those people are more likely to exist in an OS pool than a CS pool and more likely to be opponents so not share their results.
Deep learning may show the equivalent of political (or masculine and feminine). As well as encouraging group-think, that also distorts the distribution of biases and therefore the cancelling out of errors can no longer be assumed.
Human factors in defeating security often work better than exploiting software bugs. Some of the deep learning AI is designed to mimic humans as well as possible in thinking and in interfacing. I suspect that might also make them more vulnerable to meta-human-factor attacks. Again, exposure to different and diverse cultures will show a non-uniform distribution of error/bias spotting/disclosure/exploitation.
Deep learning will become harder for humans to understand as it develops and becomes more machine dependent. That will amplify the above weaknesses. Think of optical illusions that greatly distort human perception and think of similar in advanced AI deep learning. Errors or biases that are discovered will become more valuable to an opponent since they are less likely to be spotted by others, increasing their black market exploitation risk.
I have not been a programmer for over 20 years and am no security expert so my reasoning may be defective, but at least now you know what my reasoning was and can therefore spot errors in it.