acceptodds
Under review as a conference paper at ICLR 2027

AliasShare: A Plug-and-Play Module for Efficient Embedding Operations in MPC Workflows

Abstract

We propose AliasShare, a method designed to accelerate embedding operations in secure multi-party computation (MPC) for private information retrieval (PIR) in ML applications. AliasShare applies a one-to-many mapping for high-frequency tokens along with shuffling in the offline stage, reducing the need to transmit full embeddings and thereby lowering client's online upload communication overhead by ( represents embedding dimension). Meanwhile, it avoids performing costly secret embedding multiplication on the server, mitigating server computation overhead to plaintext-level. Our code is available at https://anonymous.4open.science/r/AliasShare-main-BC51.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
multi-party computation;private information retrieval

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_J4mnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes AliasShare to accelerate embedding lookup in MPC inference. A trusted model owner assigns aliases to tokens, shuffles the embedding table, and distributes secret shares to the servers and the alias mapping to the client. The client sends the alias directly to the server and each server just returns the corresponding embedding shares. The model owner periodically reshuffles the mapping and table shares to limit accumulated privacy leakage. The paper reports lower online lookup costs than several baselines and evaluates resistance to frequency-matching and substitution-decipherment attacks.

Critical strengths
  1. The idea is sensible: aliases allow direct lookup of embedding shares while preserving the original embedding values and compatibility with the subsequent MPC computation.
  2. The reported online lookup performance looks solid. For GPT-2 with 1,024 input tokens, Table 3 reports 0.036 ms of computation versus 5.12 ms for SHAFT-PIR, about a 142× speedup for this stage.
Critical weaknesses
  1. I find the technical contribution limited. Similar techniques exist in frequency-smoothing encryption[1], while secret permutations and periodic reshuffling are pretty well-studied in the original square-root ORAM construction[2] and its later adaptation to MPC[3]. These works are not cited.

  2. If I understand correctly, suppose the client needs to fetch the same token, it sends the same alias to the servers, which is a privacy leakage? Or are you doing caching?

  3. The performance comparison does not account for these differences. Direct lookup is expected to be cheaper when addresses are visible. Section 3.4 also requires the model owner to reshuffle and retransmit the entire expanded table after every K tokens, with K = 1024 in the attack evaluation. This recurring cost is outside the reported online latency formula. I would like to see amortized communication and throughput, including refreshes, before accepting the claimed efficiency advantage.

  4. I could not find an implementation of the proposed AliasShare algorithm in the released code. This limits my confidence in reproducibility.

[1] Marie-Sarah Lacharité and Kenneth G. Paterson. Frequency-Smoothing Encryption: Preventing Snapshot Attacks on Deterministically Encrypted Data. TOSC 2018.

[2] Oded Goldreich and Rafail Ostrovsky. Software Protection and Simulation on Oblivious RAMs. JACM 1996.

[3] Samee Zahur et al., Revisiting Square-Root ORAM: Efficient Random Access in Multi-Party Computation. S&P 2016.

Questions for authors
  1. What privacy guarantee holds for the complete alias-query sequence, including repeated accesses? What information is explicitly allowed to leak?
  2. What are the amortized communication and throughput when the full table is refreshed every 1,024 tokens? How does the tradeoff compare with shuffled-table ORAM or private lookup under comparable privacy requirements?
  3. Could you release the full codebase?
Reviewer_rdMDOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposed an acceleration method for MPC-based LLM inference. Its core design includes a frequency-aware alias mapping bucket and a re-shuffle protocol. The authors also evaluated their solution.

Critical strengths

None.

Critical weaknesses
  • The authors claim their method can accelerate PIR, yet they provide no formal definition of PIR. Furthermore, I do not consider secret embedding lookup to be PIR. Therefore, many comparisons are unfair.
  • There is no formal security definition.
Questions for authors
  1. What is the relationship between the attacks in Section 4.2 and the security analysis in Appendix B? Why are some established through proof, while others are not?
  2. What are the differences in the security guarantees provided by AliasShare and other solutions?
Reviewer_EG8ZOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

AliasShare proposes a faster way to look up embeddings in MPC inference. It uses alias mapping and periodic table reshuffling, allowing servers to retrieve embedding shares through local lookup. The experiments show large reductions in online communication and lookup time across several embedding tables. The main contribution is the efficiency improvement to private inference.

Critical strengths
  1. The online speedups in Tables 2 and 3 are substantial. Reducing client uploads and server-side lookup work is useful when the client has limited uplink bandwidth.
  2. The method keeps the original embeddings and the subsequent Transformer computation unchanged, which helps it fit into existing MPC workflows. Table 6 reports task accuracy after integration with PUMA and SHAFT.
Critical weaknesses
  1. Section 3.4 requires the model owner to reshuffle and resend the embedding table periodically. Section 4.2 uses K=1024 for the attack evaluation. I could not find the time and communication cost of these recurring updates alongside the online results. Even if updates are prepared in advance, they still consume compute and bandwidth. I would like to see what the method saves over repeated queries once these costs are included.
  2. Section 3 explains the model owner’s role and assumes non-colluding parties. I am still unsure whether the baselines provide the same privacy guarantees under the same trust assumptions. Please compare what the model owner must keep secret and what each server can observe across the methods. These differences matter for interpreting the speedups.
Questions for authors
  1. How long does each table refresh take, and how much data is transmitted? Including these refreshes, what are the average cost per request and the sustained throughput?
  2. Why was the reshuffling window set to K=1024? How would changing K affect the refresh cost and the defense against the evaluated attacks?
  3. Compared with GPU-DPF, SHAFT-PIR and PUMA-PIR, does AliasShare place different trust requirements on the model owner or reveal different information to the servers? If so, how should we interpret the speedup comparison?
  4. Figure 2 reports embedding/PIR latency. For one full MPC inference example, how much time is spent on embedding lookup, and what overall speedup would the faster lookup provide?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.