SETTLER: A Hardware-Assisted Secure and Efficient Transformer Inference Service

Summary

SETTLER is a hardware-assisted inference service for transformer models that uses homomorphic encryption to outsource linear layers to GPUs with cryptographic guarantees, addressing model privacy leakage in prior lightweight obfuscation schemes that rely on TEEs. The authors report inference accuracy comparable to plaintext inference and runtime speedups of 12.54-38.08x over a TEE plus HE-based matrix multiplication baseline (CCS'25), and up to 1.4x over a prior TEE-only baseline.