PLUS: The latest AI and tech news.
By Jennifer Conrad | 07.22.21   | AI can be foiled in unexpected ways. The results can be funny (flagging an ad on the side of a bus for jaywalking) or deeply offensive (tagging images of Black people as gorillas). But when AI is brought to the battlefield, a faulty system can be downright dangerous. This week Will Knight looks at efforts within the Pentagon to "red team" existing models for weaknesses and seek hidden vulnerabilities in the code and data used within the Department of Defense. The initiatives look, in particular, at thwarting adversarial attacks when a small change to an input can create a potentially catastrophic result. "A machine-learning algorithm trained to recognize certain vehicles in satellite images, for example, might also learn to associate the vehicle with a certain color of the surrounding scenery," writes Knight. "An adversary could potentially fool the AI by changing the scenery around its vehicles. With access to the training data, the adversary also might be able to plant images, such as a particular symbol, that would confuse the algorithm." When it comes to military applications of AI, the stakes couldn't be higher, and well-resourced adversaries have strong incentives to mess with military systems. Learn more about the Pentagon's efforts to make its AI systems more secure. |  | The Defense Department often looks to the tech world for new tools—but efforts to modernize can be painfully slow. | |
Post a Comment