BlackBerry's acquisition of Cylance raises eyebrows in the security community

Backdoors in machine learning have gotten little attention, but researchers have demonstrated proofs of concept for how such backdoors might work. ‘It’s possible they could add machine learning-specific backdoors of the style we proposed last year that makes it ignore their own state-sponsored malware,’ Brendan Dolan-Gavitt, an assistant professor in the computer science and engineering department at New York University [Tandon School of Engineering], tells CSO. ‘We showed that when you’re training something like a deep learning system you can teach it to recognize specific triggers and then misclassify any inputs that have that trigger,’ Dolan-Gavitt adds. ‘We haven’t looked at anti-malware systems specifically, but I think it would work.’