Kraken: Near-field and far-field high-order electromagnetic side-channel attacks against DNNs

The paper proposes Kraken, a high-order electromagnetic side-channel attack method against DNNs that can extract model parameters on GPU Tensor Cores. The authors claim this is the first parameter extraction achieved on Tensor Cores, whereas prior work targeted only general-purpose CUDA Cores. The method estimates energy consumption for the GPU architecture and carries out the attack via correlation power analysis (CPA). The authors also conduct an exploratory analysis of hyperparameter and weight leakage from LLMs in the far field, showing that GPU electromagnetic radiation can leak through glass barriers at a distance of 100 cm.