Detecting adversarial backdoor attacks on computer vision systems
Backdoor attacks on AI systems are possible by embedding a "trigger" pattern in the training data such that when observed, an unintended behavior in the system is elicited. In this work, naural-appearing triggers were inserted into camera and satellite imagery using the TrojAI framework developed by APL. After validating the effectiveness of the triggers, detectors will be developed in future work with a goal of providing a capability to ensure that computer vision systems are safe for use in critical applications.
Intern: Jeffrey Boman
Mentor: Christopher Ratto and Aurora Schmidt (REDD)