- 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
- The idea is sensible: aliases allow direct lookup of embedding shares while preserving the original embedding values and compatibility with the subsequent MPC computation.
- 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
-
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.
-
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?
-
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.
-
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
- What privacy guarantee holds for the complete alias-query sequence, including repeated accesses? What information is explicitly allowed to leak?
- 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?
- Could you release the full codebase?