Toward Double-Layer Data Privacy in Communication-Efficient Hierarchical Federated Learning: A Client Sampling Approach
Bibliographic record
Abstract
Federated Learning (FL) is a promising learning paradigm that allows for training a shared model by coordinating multiple distributed devices, namely, clients, without exposing their raw data. To mitigate excessive communication overhead and enhance practicality, a variant known as Hierarchical FL (HFL) has been introduced, which integrates edge servers between the cloud server and clients. In HFL, the number of potential clients is typically large, making full client participation impractical due to various resource constraints. As a result, it is essential to develop a sampling strategy that effectively selects suitable clients for federated optimization. While several methods have been proposed to protect the privacy of communicated models, we argue that the outcomes of client sampling are closely tied to the local data of clients, thereby raising privacy concerns, like the risk of differential attacks. Motivated by this, we propose a Two-step Privacy-Preserving client Sampling framework (TPPS) designed to protect against both attacks on communicated models and potential vulnerabilities in client sampling outcomes. Initially, we consider the diverse privacy requirements of clients by presenting a double-layer noise mechanism. We then conduct a thorough analysis of the impact of this noise mechanism, proposing a novel client sampling strategy that seeks to balance the trade-off between privacy and training performance. The insight lies in maintaining a real-time sampling probability for each client, which can be acutely tuned based on personalized privacy needs and previous training feedback. We provably show that TPPS achieves local differential privacy, a bounded sampling regret, and a privacy-related convergence rate. Furthermore, we conduct extensive simulations based on open datasets, showing the robustness and applicability of TPPS in enhancing privacy while optimizing HFL performance.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.009 | 0.026 |
| Meta-epidemiology (narrow) | 0.001 | 0.001 |
| Meta-epidemiology (broad) | 0.002 | 0.001 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.001 | 0.003 |
| Scholarly communication | 0.002 | 0.005 |
| Open science | 0.004 | 0.006 |
| Research integrity | 0.002 | 0.004 |
| Insufficient payload (model declined to judge) | 0.002 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".