An Army research office and an intelligence research agency are looking for proposals on technology and methods to detect trojans in artificial intelligence.
Trojans in Artificial Intelligence, or TrojAI, is billed as a 2-year program with several awardees working together to develop and create software that automatically inspects an AI and predicts if it has a backdoor virus.
The recently released broad agency announcement comes from the U.S. Army Research Office and the Intelligence Advanced Research Projects Activity.
As outlined in the BAA, a trojan attack could be used to change an AI to give a false response. AI is trained on data and then unleashed to the world to operate on new data. In the example highlighted in the announcement, an AI can be trained on images of traffic signs and learn what various signs look like. The AI could then be used in a driverless car.
An adversary could disrupt the training pipeline and insert trojan behaviors into the AI, causing the car to ignore the stop sign as the AI would incorrectly see it as a speed limit sign. To prevent that from happening, the TrojAI program wants to be able to detect if an AI has a trojan to ensure it can be securely implemented.