One voice hack won’t fit all: New AI method breaks mass attacks
What Happened
Researchers Qiang Zeng and Ying Meng at George Mason University have developed DivNet, a technique designed to prevent the same adversarial audio attack from working across large numbers of voice-assistant devices. The problem is that manufacturers can deploy identical AI models across millions of devices. DivNet instead creates many slightly different versions of the underlying model by selectively modifying neural-network weights that have little impact on legitimate speech recognition but influence malicious attacks.
Key Takeaways
Tests on established speech-recognition systems found that DivNet substantially reduced attack success with only minimal loss in normal recognition accuracy. It also added no processing delay during speech recognition, while individual model variants could be generated in milliseconds after the initial analysis.