acceptodds
Under review as a conference paper at ICLR 2027

Send the Brain, Keep the Body: Hybrid Private Transformer Inference with Minimal Secret Control

Abstract

Hybrid private inference combines the computational strengths of homomorphic encryption (HE) and secure multiparty computation (MPC), but repeatedly converting intermediate states between them incurs substantial communication. Prior work optimizes operator partitioning and evaluation, but still organizes HE–MPC execution around transfers of intermediate states. We observe that the secret decisions involved in nonlinear evaluation can be separated from the high-dimensional numerical state. Based on this observation, we introduce SHRINK, a hybrid Transformer inference architecture with decoupled data and control paths. SHRINK sends MPC only decision-critical information, while HE continues the main computation on retained ciphertexts. MPC generates secret controls that HE applies to existing ciphertexts, enabling continued HE evaluation across nonlinear boundaries. We exploit Transformer structure to design compact control protocols for GELU, attention, and LayerNorm. We further tailor ciphertext representations and feedback to the HE operations that use these controls, and reuse control computations across data blocks, reducing communication and redundant homomorphic computation. On BERT-base and BERT-large, SHRINK reduces communication by 8.4–1048.6 compared with hybrid inference baselines, while maintaining accuracy close to that of plaintext inference without retraining.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Private Transformer Inference, Homomorphic Encryption, Secure Multi-Party Computation, Privacy-Preserving Machine Learning

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 2 reviews, above 9% of the others.
Reviewer_xNWKOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The work targets on reducing the communication and computational overheads during converting intermediate states between HE and MPC in secure transformer inference. The paper proposed several control protocols for GELU, Softmax and Layernorm. The paper conducted experiments on BERT-base and BERT-large to show its improvements. The motivation makes sense, but the technical and experimental details are a bit difficult to follow.

Critical strengths

The paper includes many figures to explain the technical details. The figures are nice.

The paper considers baselines from both HE-only methods and HE+MPC methods.

Critical weaknesses
  1. The work only considers 3 functions in transformer and only tests the proposed methods in BERT style models. It’s unclear that whether the proposed method can be applied to other transformer models like Llama, and whether the proposed GELU control can be used in other activations like Silu.

  2. The implementation details are missing. The paper doesn’t provide codebase and also doesn’t clear describe the library, HE parameters, security levels and other reproducibility details.

  3. The paper misses some important related works and baselines in secure transformer inference like MOAI[1], ARION[2], PUMA[3] etc

  4. The description is unclear. For example, Table 2, it doesn’t introduce what is P(x). In section 4.1, there are many notations in Figure 2 such as PR(x), QR(x), EP(), but authors don’t explain them.

[1] MOAI: Module-optimizing architecture for non-interactive secure transformer inference. ICLR 2026

[2] ARION: Attention-optimized transformer inference on encrypted data[J]. Cryptology ePrint Archive, 2025.

[3] Puma: Secure inference of llama-7b in five minutes[J]. Security and Safety, 2025, 4: 2025014.

Questions for authors
  1. How to decide the parameters like K, R in the proposed method? How will them affect the efficiency and accuracy?

  2. In Table 3, does the latency of SHRINK considers the communication latency? What network setting is used in this evaluation?

  3. What packing methods and matrix multiplication algorithms are used in SHRINK?

Reviewer_enraOpenReview ↗
Rating
2
Confidence
4
Informativeness
2(+1)
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents SHRINK, a hybrid HE–MPC framework for private Transformer inference that retains intermediate states in HE and uses MPC to generate compact, input-dependent controls. It develops protocols for GELU, attention, and LayerNorm, together with projection-based ciphertext feedback compression and computation scheduling. In particular, its dense-attention protocol reduces communication from quadratic to linear in sequence length while retaining quadratic attention computation. The paper provides interface-level security proofs and conditional numerical error bounds. Experiments on BERT-base and BERT-large across five GLUE tasks report 8.4× and 11.0× lower communication than BLB, respectively, with near-plaintext accuracy without retraining. Ablations and sequence-length experiments examine the contributions of individual mechanisms and communication scaling.

Critical strengths
  • SHRINK retains the attention scores in HE and exchanges only row-level statistics and controls, reducing attention communication from quadratic to linear in sequence length while preserving dense computation. The reported 8.4×/11.0× communication reductions over BLB on BERT-base/large。

  • The projected-feedback construction preserves the ciphertext obtained after the prescribed projection, and Figure 9 demonstrates why feedback representation matters: HE-based LayerNorm refinement alone increases communication by approximately 9%, whereas combining it with projected feedback reduces communication by 33.6%.

Critical weaknesses
  • The numerical guarantees remain conditional, with limited evidence that their assumptions hold throughout execution. The attention approximation assumes a bounded score domain, while GELU channel allocation depends on LayerNorm output norms and upstream arithmetic-error budgets. The paper does not clearly establish that these conditions remain satisfied across the complete network. Table 10’s local-error measurements use one instrumented MRPC execution per model and cover only attention-output LayerNorm.

  • The security theorem does not clearly establish the full private-inference guarantee. Theorem B.6 compares the concrete protocol with an ideal-interface execution that retains the surrounding HE computation and local state. This establishes a useful interface-level result, but the paper does not clearly connect that reference execution to an end-to-end functionality exposing only the prescribed output and permitted leakage.

Questions for authors

see above.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.