Pith. sign in

REVIEW 2 cited by

Privacy Amplification via Compression: Achieving the Optimal Privacy-Accuracy-Communication Trade-off in Distributed Mean Estimation

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2304.01541 v1 pith:ZC2INRYI submitted 2023-04-04 stat.ML cs.CRcs.LG

Privacy Amplification via Compression: Achieving the Optimal Privacy-Accuracy-Communication Trade-off in Distributed Mean Estimation

classification stat.ML cs.CRcs.LG
keywords varepsilonleftrightprivacybitsclientcompressionestimation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved
0 comments
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Privacy and communication constraints are two major bottlenecks in federated learning (FL) and analytics (FA). We study the optimal accuracy of mean and frequency estimation (canonical models for FL and FA respectively) under joint communication and $(\varepsilon, \delta)$-differential privacy (DP) constraints. We show that in order to achieve the optimal error under $(\varepsilon, \delta)$-DP, it is sufficient for each client to send $\Theta\left( n \min\left(\varepsilon, \varepsilon^2\right)\right)$ bits for FL and $\Theta\left(\log\left( n\min\left(\varepsilon, \varepsilon^2\right) \right)\right)$ bits for FA to the server, where $n$ is the number of participating clients. Without compression, each client needs $O(d)$ bits and $\log d$ bits for the mean and frequency estimation problems respectively (where $d$ corresponds to the number of trainable parameters in FL or the domain size in FA), which means that we can get significant savings in the regime $ n \min\left(\varepsilon, \varepsilon^2\right) = o(d)$, which is often the relevant regime in practice. Our algorithms leverage compression for privacy amplification: when each client communicates only partial information about its sample, we show that privacy can be amplified by randomly selecting the part contributed by each client.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Shuffling-Aware Optimization for Private Vector Mean Estimation

    cs.LG 2026-04 unverdicted novelty 7.0

    Using the shuffle index, the authors formulate and solve an optimization problem for post-shuffle minimax-optimal unbiased mean estimation, yielding an asymptotically optimal mechanism whose privacy-utility tradeoff a...

  2. FedRP: A Communication-Efficient Approach for Differentially Private Federated Learning Using Random Projection

    cs.LG 2025-09 reject novelty 5.0

    FedRP claims to preserve FedAvg-level accuracy while sending only a few numbers per client per round and providing an (epsilon, delta)-DP guarantee.