Security Review: AI
Although thanks to Sarah Connor from The Terminator, 1997 wasn’t quite the day machines went out of control, autonomous systems become more and more integrated into our lives. Although AI might take many years to get developed to the level of human intelligence, it is also possible that breakthrough is just around the corner. Therefore, in order to not be taken by surprise it is important to start addressing security of the autonomous systems on the broad level.
In this article I will address several concerns and possible ways to deal with them, as there are multiple assets at stake – human lives, material things, in fact, almost everything of reach and a little beyond.
I would categorize aspects of AI into two major categories – AI on mobile devices and AI on stationary devices, of which I will focus more on mobile, as, from the first glance, they seem to be the most dangerous, as can cause direct physical harm.
Who might be potential adversaries in the context of AI systems? Here are some of them: Creator of the devil machinery, outside person willing to exploit and manipulate the system, by giving it incentives. And apparently, AI system itself – even without the intention of the creator, some unknown processes might happen in neurons of the system.
There can be numerous judicial and technological means that can be implemented to reduce possible negative outcomes. I am not a big fan of restrictions, as it will drag development of the technology. Nevertheless, here are they:
Obvously, protection from unauthorized access and intrusion detection, using one of the common methods to prevent from accessing and manipulating hardware directly.
Here are more related specifically to AI:
Restriction on the AI algorithm – do not allow algorithm which seemed to work, but could not be fully understood take care of the children, for example.
Restrictions on incentives for mobile robots – it is important to carefully think about what stimulus can be left and what should be eliminated, so that system doesn’t create a danger striving to reach for that incentive. Note, it is possibly, will not be possible to make it always happy, as it might restrict its willingness to learn.
Allow only pre-learned machines to go wander on their own, thus, disabling learning capabilities.
Requirement of restricted weight/power – humans should be able to deal with them.
Although the cars that drive themselves will likely be safer, it is not so clear with fully developed learning AI system implanted in flexible mobile machinery, unless necessary precautions are implemented.
Additionally, there are numerous other questions, including ethical coming with further development of AI, such as whether it can be considered a slavery, for example. Regulations on Artificial intelligence systems are inevitable, and users and developers should be thinking of them and be prepared for them.