The Increasing Importance of AI Security
05/06/2021
Over the previous years, we have seen strong shifts towards using AI to solve security problems both in the digital world and the real world.
Such solutions range from automated log analysis to cameras with facial recognition features in public places. And since the cost of any AI system - once the problem is solvable by AI - is negligible when compared to human labor, this trend is very likely to continue over the coming years.
The risks of AI automation
(Un)fortunately, AI systems are not foolproof.
While media portrayals like to depict them as sages, their actual behavior is closer to 4 year old children with vast capacities. They lack all awareness of circumstances, social contexts, history and social norms. As such, any early-stage AI that solves a problem usually ends up providing a solution that is correct but useless. If you tell a 4 year old to clean their room without any further guidelines in place, you will almost certainly find all clothes and toys either in a pile or under the bed. Similarly, if you ask an AI to identify fraud in logs, it will almost certainly start singling out factors that are correlated but not usable in modern society (e.g. names written in a specific alphabet or IPs belonging to certain countries).
At the same time, the very core systems that AI relies on - from back-propagation to convolutional layers - are all approximations of human intelligence. While they often produce very good results, this difference means that inputs that look, read or sound different to human observers and AIs can be easily created.
And beyond that lies the minefield of the security of AIs themselves. Can trained AIs leak information about the data they were trained on? How do you secure the supply chain of your training data?
The future
Of course, all of these questions are being worked on - both on the offense and defense - by countless specialists and researchers. But as the field develops, we are almost certain to see multiple high level incidents related to AI security during the next 5-10 years.
We at Reflare have been involved in some of this research together with Deloitte, and our CEO, Paul S. Ziegler, will be presenting the results at an upcoming Black Hat event.
Naturally, we will also continue to update you on relevant findings here.
Copyright © 2021 Reflare Ltd, All rights reserved.
SME- Retired (1/31/2024)
4yAgree. Thanks for sharing.