OptiPrime: Optimizing Private Inference through Protocol-Hardware Co-design
Jiangrui Yu, Ye Yu, Si Chen, Chenqi Lin, Wenxuan Zeng, Wenxuan Zeng, Junfeng Fan, Mingyu Gao, et al.
cs.AR · cs.CR · cs.LG
Quick verdict
Cuts private-inference latency up to 5.7x vs Cheetah via protocol-FPGA co-design; strong real-hardware evaluation but closed platform limits reproducibility.
Summary
OptiPrime addresses a deployment gap in private DNN inference: commercial HE accelerators speed up individual homomorphic operations by orders of magnitude (444x for ciphertext-plaintext multiplication, Fig. 1b), yet applied to the Cheetah HE-MPC protocol they yield only 1.37x end-to-end speedup because transmitting output ciphertexts over the network dominates. The co-design has three parts. OptiEncode encodes tensors along the input-channel dimension so automorphism sums remove dummy coefficients, cutting output ciphertexts by up to 128x; a baby-step-giant-step reformulation keeps automorphism count near Cheetah's zero (Table I: 10 vs 1280 without BSGS for a ResNet-18 layer). OptiComp exploits >99% sparsity of channel-encoded weight plaintexts with a segment format and a 0.8k-LUT decompression unit (0.7% of the U55C FPGA). OptiFlow is an output-stationary dataflow that recomputes baby-step ciphertexts on the fly to avoid HBM thrashing. On ImageNet CNNs (W8A8), ResNet-18/50 drop to 2.9 s and 14.6 s, up to 5.7x faster than Cheetah on CPUs and 4.2x with the accelerator, at unchanged accuracy. Evaluation spans LAN/WAN emulation, bandwidth/RTT/HE-parameter sweeps, an ARK simulator, and BERT-base linear layers. Main limitations: a closed commercial platform, no released code, emulated networks, and simulator-based ARK comparisons.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Applying a commercial HE accelerator to Cheetah yields only 1.37x end-to-end speedup despite 444x per-operation gains, because output-ciphertext network transmission dominates (Fig. 1).
- OptiEncode's channel encoding with automorphism cleaning reduces output ciphertexts by up to 128x vs Cheetah, and its BSGS reformulation cuts automorphisms by >128x vs the naive version (Table I: 1280 to 10 for a ResNet-18 layer).
- OptiComp reduces weight-plaintext memory traffic by 10x (abstract) via lossless sparse-segment compression with a 0.7%-area decompression unit, and OptiFlow cuts total memory access to 0.03x-0.21x of baseline (Fig. 17).
- End-to-end, OptiPrime is up to 5.7x faster than Cheetah on CPUs and 4.2x faster with an accelerator, reaching 2.9 s (ResNet-18) and 14.6 s (ResNet-50) on ImageNet at unchanged accuracy (Table III).
- The design generalizes across HE parameters (WAN speedup grows to 12.9-13.6x at N=32768, Fig. 22), network conditions (Fig. 20), an ARK simulator (Fig. 18), and BERT-base linear layers vs Iron and BumbleBee (Fig. 19).
Strengths
Sharp motivating measurement
Fig. 1b-d quantifies the deployment gap: 444x/182x per-operation speedups translate to only 1.37x end-to-end on Cheetah, with output-ciphertext NetIO identified as the bottleneck. The paper is built on a measured failure mode, not an assumption.
Formally specified, checkable protocol
The pi_x/pi_w encoding definitions, the identity sigma(.,N+1)=sigma^2(.,N/2+1), and the BSGS decomposition (Eq. 1) make Table I's counts independently verifiable: 2048 CPMults, 10 automorphisms, 8 output ciphertexts for a ResNet-18 layer vs Cheetah's 2048/0/256 and no-BSGS 1280 automorphisms.
Real-hardware implementation, not simulation
Unusually complete evaluation chain
Honest scoping
Weaknesses
Reproducibility bottleneck
10x compression claim not tabulated
Emulated and adapted baselines
Opaque headline range in Sec. V-B
Writing defects
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
OptiPrime — Structured Preliminary Review
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What problem does OptiPrime solve in private AI inference?
In HE-MPC private inference, hardware accelerators make the homomorphic math fast, but many output ciphertexts must still be sent over the network after every linear layer. The paper measures that a commercial FPGA accelerator applied to Cheetah speeds up individual HE operations by 444x yet the whole inference only 1.37x, because network transfer of output ciphertexts dominates. OptiPrime's new encoding cuts those output ciphertexts by up to 128x, so accelerator gains translate into end-to-end speedups.
How fast is private inference with OptiPrime?
On ImageNet with W8A8 quantization, ResNet-18 drops to 2.9 seconds and ResNet-50 to 14.6 seconds with the FPGA accelerator. Overall, OptiPrime is up to 5.7x faster than the Cheetah protocol on CPUs and up to 4.2x faster with an accelerator, at unchanged prediction accuracy (for example, ResNet-50: 75.93% vs 76.14% FP32 top-1).
Does the new protocol or the compression change model accuracy?
No. OptiEncode changes only how coefficients are laid out and keeps results exact; OptiComp and OptiFlow are server-side memory optimizations. Measured ImageNet top-1 accuracy under W8A8 matches prior protocols exactly (Table III); the only loss comes from quantization itself, and the accumulator stays below 19.3 bits versus a 21-bit plaintext modulus.
Is OptiPrime's code available?
The paper does not announce a code release. It builds on the commercial LattiSense FPGA accelerator platform, and the comparison against the ARK accelerator uses a simulator the authors built from published parameters, since ARK's hardware and software are not public.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The abstract is truncated mid-sentence in the provided text; headline speedups were cross-checked against Sec. I and the conclusion.
- Figures are flattened text; per-bar values in Figs. 13-22 are scrambled in extraction, so range attributions rely on the prose rather than chart values.
- Bandwidth units ('384MBps'/'44MBps') are quoted as written; MB/s vs Mb/s ambiguity is unresolved.
- The 'wireless' phrasing in the abstract and Sec. I likely means wireline/wide-area transmission; treated here as a wording issue, not a technical claim.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.