Boffins trick AI model into giving up its secrets

NC State researchers developed "TPUXtract", a side‑channel electromagnetic technique that extracts neural network hyperparameters from Google Edge TPUs (used in Coral Dev Board and similar accelerators), allowing attackers to recreate models with high fidelity (reported 99.91% accuracy). The attack assumes physical access and EM measurement equipment, takes roughly three hours per layer in tests on models like MobileNet V3, Inception V3, and ResNet-50, and highlights risks for devices without memory encryption.