acceptodds
Under review as a conference paper at ICLR 2027

Shape-Selective Correlation Conversion for Public-Shape Private Transformer Inference

Abstract

Private inference systems typically prepare correlations for a fixed tensor graph, whereas token pruning and elastic networks may reveal a public tensor shape only after preprocessing has been distributed. Preparing every supported shape duplicates correlations, while generating and delivering preprocessing after disclosure delays online execution. We introduce shape-selective correlation conversion (SSCC), which locally specializes distributed preprocessing while preserving the fixed-shape online protocol. For retained bilinear correlations, we identify the missing contraction term and characterize the exact minimum additional ring elements required by zero-interaction share-linear decoders over finite chain rings. We instantiate SSCC for nested attention capacities and elastic SwiGLU widths and compose both axes in one Transformer block without a capacity-by-width preprocessing family. For public shape selection and one-use preprocessing, conditional completion lifts fixed-shape security under branch-isolation and joint-pseudorandomness assumptions. Across nine capacity-width pairs, selected and separately generated fixed-shape outputs agree exactly and incur identical online communication. Factorized preprocessing reduces per-party storage by 50.39% relative to a capacity-deduplicated fixed-shape baseline, and a synthetic 12-block chain checks cross-block output and communication consistency. A BERT-base case study on the Microsoft Research Paraphrase Corpus matches the unpruned accuracy of 86.52%. Across 20 co-located repetitions per method and width excluding delivery, the observed median interval from public width disclosure to output is 16.3% to 20.7% lower than for generating fixed-shape preprocessing after disclosure.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
private transformer inference, secure multi-party computation, function secret sharing, token pruning, communication-efficient inference

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_eWNUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the preprocessing problem in FSS-based secure inference, a well-chosen angle because while communication and latency in the online phase of FSS-based secure inference are already well optimized, research on the preprocessing phase is limited. This paper primarily focuses on how to obtain the required preprocessing from the distributed materials without increasing interaction or altering the online protocol, when the token capacity or FFN width is only determined after the preprocessing material is distributed. The authors propose Shape-Selective Correlation Conversion (SSCC), classifying materials into projective correlations, bilinear correlations requiring correction, and shape-specific correlations—a valuable observation. In a zero-interaction model with linear decoding of correlation shares, the authors prove matching upper and lower bounds on the number of additional scalar correlations and discuss the security of retaining all preprocessing material through conditional completion.

Critical strengths
  1. The authors have effectively identified the bottleneck in the preprocessing stage. Compatibility between dynamic shapes and dealer-based preprocessing mechanisms is not easy, as preprocessing relies on fixed shapes, and the actual execution dimensions may only be determined in subsequent calculations. This paper proposes a solution to this problem.

  2. Maintaining the existing online protocol is another point I appreciate, as it means the method can potentially be directly integrated into existing SIGMA/FuseFSS workflows.

  3. The combined processing of capacity and width is also a significant contribution. This combined processing organizes the material into common parts, capacity-indexed parts, and width-indexed parts, which is clear and avoids duplicating a complete set of preprocessing material for each combination.

Critical weaknesses
  1. The source of the cost advantage is not yet clear. The paper primarily compares two scenarios: "pre-preparing all shapes" and "generating only after shapes are determined." However, in reality, if unused fixed-shape materials can be safely reserved for subsequent requests, the system may only need to replenish the actually consumed portion. Always executing at the maximum shape is another viable option. The authors need to explain the feasibility of these options, what additional costs they incur, and under what conditions SSCC is more cost-effective than other scenarios.

  2. The description of the Softmax method is inconsistent with storage statistics. The paper states in one place that some Softmax materials can be shared across different shapes, while in another place it prepares and bills all Softmax materials separately for different shapes. This seems contradictory to me, and I am unclear which implementation was actually used in the experiments. The authors should clarify which materials are shared and which are generated separately, allowing readers to determine the source of the storage savings.

  3. The evidence for actual operational benefits is not complete. In the joint block experiment, online execution of SSCC was actually slower than fixed-shape execution, but the reason is not explained. The full MRPC request uses a separate HE/MPC system, which doesn't directly verify the actual benefits of SSCC.

  4. End-to-end SSCC execution wasn't performed on the full trained model. I'm curious about the benefits of using SSCC on a full model.

Questions for authors
  1. Does Appendix D.1 describe the actual running implementation or an optimization that has not yet been implemented?

  2. Considering that Appendix F.1 exposes remapping information before selection, can unused fixed-shape instances be safely reserved for subsequent requests? If so, under what conditions would SSCC truly outperform such a preprocessing scheme?

  3. Where does the performance gap as a function of width in Table 7 come from?

  4. Can end-to-end SSCC be demonstrated on a trained model, even a small one?

Reviewer_yA5BOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper targets private Transformer inference and material is usually prepared for one fixed shape, but token pruning and elastic widths fix the shape mid-request. The dealer sends material for the largest shape once. After the shape becomes public, parties convert it locally, with no extra communication. Matrix products also need correction material to cancel terms from dropped coordinates. Outputs and communication match the fixed-shape protocol across 9 settings. Storage is 50.39% lower. BERT-base keeps 86.52% accuracy on MRPC. Latency drops 16.3–20.7%, excluding delivery time.

Critical strengths
  • The authors identify why simply cropping the max-shape material fails: a shortened summed dimension leaves terms from dropped coordinates in the product, achieveing good preformance on storage and accuracy.

  • The paper lets the dealer prepare once and the parties specialize locally after the shape is chosen. The online protocol, its communication cost and its outputs stay identical to the fixed-shape version, and that match is checked across 9 configurations and a 12-block chain.

Critical weaknesses
  • The framework's security rests on extra, non-standard assumptions. The fixed-shape security guarantee only carries over to SSCC under “branch isolation” and “joint pseudorandomness.” From what I’ve read, it isn’t clear whether real implementations satisfy these, or what an attacker gains if they don’t.

  • Lack of novelty. The core technique may be an incremental extension of known preprocessing methods. Restricting maximum-size correlated randomness to a subset of coordinates and adding correction terms is close to standard practice in secret-sharing MPC.

  • Lack of baselines. the comparison set is narrow, and two natural baselines appear to be missing. The abstract and introduction mention only two comparisons. For storage, it’s a fixed-shape baseline that shares material across capacities. For latency, it’s generating fixed-shape material after the shape is disclosed. Several obvious alternatives seem absent: Zhang, Yancheng, et al. "Cipherprune: Efficient and scalable private transformer inference." International Conference on Learning Representations. Vol. 2025. 2025. Cai, Yifei, et al. "SecDTD: Dynamic Token Drop for Secure Transformers Inference." arXiv preprint arXiv:2603.13670 (2026).

Questions for authors

Can you state “branch isolation” and “joint pseudorandomness” formally, explain why your implementation satisfies them, and say what leaks if they fail?

How does SSCC differ from directly applying known local conversion or subset-restriction techniques for Beaver-style correlations?

Reviewer_3m4yOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies private inference when one-use preprocessing must be distributed before the public tensor shape is known. It proposes shape-selective correlation conversion (SSCC), combining coordinate projection, corrections for shortened bilinear contractions, and shape-specific nonlinear preprocessing while preserving the selected fixed-shape online protocol. The paper characterizes the minimum explicit augmentation within a zero-interaction share-linear decoder model and instantiates the construction for attention capacities and elastic FFN widths. Experiments demonstrate selected-versus-fixed-shape output agreement and a 50.39% preprocessing storage reduction against a capacity-deduplicated baseline. Validation also includes trained FFN layers, a synthetic 12-block chain, and a separate HE/MPC BERT case study.

Critical strengths

1、The paper characterizes the correction needed when shape selection changes a contraction dimension while preserving output-mask linkage. Its attention and FFN constructions attain the stated bounds. 2、Experiments on trained FFN layers, nine joint-block configurations, and cross-block composition demonstrate agreement with fixed-shape execution and substantial storage savings against the reported baselines.

Critical weaknesses

1、Insufficiently isolated storage gains. The capacity-deduplicated baseline still stores nine FFN shares. A controlled baseline storing one maximum-row FFN per width and cropping token rows would help separate ordinary projection savings from the additional benefit of contraction-aware conversion.

2、Limited deployment evidence. The composed SSCC chain is synthetic, while the full-model MRPC study uses a different backend. Overhead relative to pre-distributed, projection-optimized preprocessing remains unclear because Table 7 uses fixed execution order and reused preprocessing. The dynamic 32/64 policy also lacks a cost comparison against always using capacity 64, which achieves the same reported accuracy and F1.

Questions for authors

1、What storage and post-disclosure request-time savings remain against a baseline that stores one maximum-row fixed-width FFN record per width and performs only row projection after capacity selection? Please use the same mask interfaces, include the associated remaps, and separate these gains from ordinary row-wise deduplication. If this baseline is not feasible, what specific correlation or interface prevents it? 2、In Appendix J, does the revealed valid-token count K refer to the original non-padding length or to a quantity produced by private computation in the first block? When does it become available, and is the intended leakage K itself or only the selected capacity? What average computation or communication benefit does the public 32/64 policy provide over always using capacity 64 at the same reported task quality? 3、Theorem 2 requires exact conditional completion in the uniform-tape hybrid, whereas Appendix F invokes computational completion and branch-key privacy. Can you state the computational version of the lifting theorem explicitly and identify the additional distinguishing-advantage term? Please also clarify which properties of the enriched fixed-shape view follow from the underlying primitives and which remain additional implementation assumptions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.