GMU researchers propose DivNet to diversify AI models to thwart voice-assistant attacks

Researchers describe DivNet as generating numerous slightly different AI models to prevent attackers from reusing the same malicious audio across devices. The approach aims to maintain accurate speech recognition while altering responses to adversarial recordings. Initial tests indicate fewer successful attacks and minimal impact on everyday use, with milliseconds to generate a new variant.

A key practical aspect is cost and deployment: training AI models remains expensive, and DivNet’s diversity could reduce repeated attack success across devices without sacrificing performance. However, the evidence comes from tests on established systems and may not apply to every device or environment, so results should be interpreted with caution.