MIT CSAIL researchers devised a way to maintain an AI model’s accuracy while ensuring attackers can’t extract sensitive information used to train it. The approach is computationally efficient, reducing a longstanding tradeoff between accuracy and privacy.