BadNets: Backdooring Attacks on Deep Neural Network Training (and What to Do about Them)

With the tremendous success of deep learning, modern artificial intelligence technology hold tremendous potential to transform human society for the better. Yet, there are emerging concerns about how this technology might be misused, either inadvertently or by malicious adversaries. This talk will provide an overview of the new security vulnerabilities in deep neural networks and specifically how they can be stealthily "backdoored" during training such that they misbehave on certain special inputs. We refer to such maliciously modified networks are referred to as Backdoored Neural Networks of BadNets. These vulnerabilities can be exploited by attackers to perpetrate harm, for instance, causing deep learning based traffic sign detectors to misclassify stop signs with special stickers stuck on them. Emerging solutions to detect and mitigate backdoor attacks, including the presenter's recent work on "fine-pruning" defenses will be discussed. Finally, implications of adversarial attacks on AI on the use of deep learning in other domains, including 5G wireless and medical informatics, will be discussed.