Researchers at Google and Atlas Computing have developed a method to extract the architecture of AI models running on specific chips by analyzing electromagnetic data. The technique, which was tested with 99.91% accuracy, involves comparing the electromagnetic field data from an unknown model to data captured while other AI models ran on the same chip. This could potentially allow attackers to reverse-engineer AI models used in smartphones and other edge devices, highlighting the need for physical security measures to protect against such attacks.
https://gizmodo.com/how-to-steal-an-ai-model-without-actually-hacking-anything-2000542423