Bifrost: Hybrid TEE-FHE Inference for Privacy-Preserving Transformer and LLM Serving
This work addresses the confidentiality problem of cloud-based LLM inference for users with sensitive data, offering a practical hybrid approach that significantly reduces overhead compared to pure FHE.
Bifrost proposes a hybrid TEE-FHE architecture for privacy-preserving LLM inference, using FHE only for linear layers on accelerators while keeping non-linear and state operations in a CPU TEE. It achieves 9.25-9.91x latency reduction on GPT-2 and LLaMA 3, and 14.6-53.4x TTFT reduction on smaller models.
Cloud-hosted transformer and large language model (LLM) inference creates a direct confidentiality problem: user prompts may contain sensitive code, business data, personal information, or regulated documents, yet remote serving exposes intermediate state to the cloud software stack and accelerator runtime. Fully homomorphic encryption (FHE) keeps accelerator-side execution ciphertext-only, but end-to-end LLM inference remains expensive because linear layers are interleaved with non-linear, cache-state, and refresh-sensitive operators. CPU trusted execution environments (TEEs) can execute those operators natively, but a CPU TEE alone does not define how an untrusted accelerator should participate. We present Bifrost, a hybrid TEE-FHE serving architecture in which secrets are provisioned only to an attested CPU TEE, while the accelerator, device memory, driver/runtime stack, and host software remain outside the trusted computing base. Bifrost uses FHE as a secure delegation mechanism for projection and feed-forward linear layers on accelerator-backed CKKS, while non-linear operators, attention-side control logic, KV-state transitions, and decrypt-then-encrypt refresh execute inside the CPU TEE. Bifrost+ further applies a prefill/decode split: prompt-side KV state is built inside the CPU TEE, and only decode-side state enters the hybrid ciphertext path. In an estimator-style comparison matching Euston's methodology, Bifrost reduces projected latency by 9.25x on GPT-2 (1.5B) and 9.91x on LLaMA 3 (8B). In direct CKKS/FHE deployments, Bifrost+ reduces TTFT by 14.6-45.8x on GPT-2 (124M) and 15.3-53.4x on Qwen3 (0.6B). The systems lesson is selective encrypted execution: use FHE only where ciphertext-only accelerator delegation is required, and keep non-linear, refresh, and prompt-side work inside the CPU TEE.