With 'TPUXtract,' Attackers Can Steal Orgs' AI Models

Researchers demonstrated "TPUXtract," an EM side-channel technique that can reconstruct CNN hyperparameters with 99.91% accuracy from a Google Edge TPU by probing chip emissions and matching signals against simulated templates; the method requires specialized, costly equipment and physical access but poses significant IP-theft and attack-surface risks, and the paper proposes mitigations such as adding noisy or randomized operations during inference.