Nights Watch

Nights Watch, an Artificial Intelligence (AI) security capability, demonstrated above Technical Readiness Level (TRL) 3 capability at xTechScalable 2.0 competition finals identifying adversarial AI attacks in data and models. Often Army Intel Soldiers are tasked with real-time situational awareness and knowing the exact position of the enemy forces. Soldiers are increasingly using image classifiers to alert when military vehicles enter the field of view of various cameras. As AI solutions make their way into battlefield operations, Soldiers will be ôon the loopö reviewing AI enemy image indicators as opposed to ôin the loopö reviewing every second of video and every image. This can lead to increased risk of Soldiers only seeing low confidence indicators. An adversarial AI attack that causes incorrect, high confidence predictions is the biggest threat to a Soldier ôon the loopö scenario.??ááNT Concepts developed Nights Watch with a team of cross-functional data scientists and developers who have successfully developed adversarial AI attacks for the DoD over the past four years. Our team is highly qualified to develop defenses to these attacks. Our team understands the human imperceptible signatures that the attacks leave, how they impact the modelĺs internal functions, and how to recover original predictions prior to the influence of an adversarial AI attack.??ááNights Watch wraps existing inference systems, making it easy to deploy and easy for Soldiers to monitor. When computers became ubiquitous and interconnected it was proved to the world that they could be hacked and afterwards antivirus became pervasive. Now, ML models have been proven vulnerable to adversarial AI, and Nights Watch is the capability Soldiers need to protect critical models, their predictions, and ultimately decisions made based upon those predictions.?á