Communication-Efficient and Privacy-Preserving Federated Learning via Joint Knowledge Distillation and Differential Privacy in Bandwidth-Constrained Networks
Why this work is in the frame
A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.
Bibliographic record
Abstract
The development of high-quality deep learning models demands the transfer of user data from edge devices, where it originates, to centralized servers. This central training approach has scalability limitations and poses privacy risks to private data. Federated Learning (FL) is a distributed training framework that empowers physical smart systems devices to collaboratively learn a task without sharing private training data with a central server. However, FL introduces new challenges to Beyond 5G (B5G) networks, such as communication overhead, system heterogeneity, and privacy concerns, as the exchange of model updates may still lead to data leakage. This paper explores the communication overhead and privacy risks facing the implementation of FL and presents an algorithm that encompasses Knowledge Distillation (KD) and Differential Privacy (DP) techniques to address these challenges in FL. We compare the operational flow and network model of model-based and model-agnostic (KD-based) FL algorithms that enable customizing per-client model architecture to accommodate heterogeneous and constrained system resources. Our experiments show that KD-based FL algorithms are able to exceed local accuracy and achieve comparable accuracy to central training. Additionally, we show that applying DP to KD-based FL significantly damages its utility, leading to up to 70% accuracy loss for a privacy budget <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\epsilon \leq 10$</tex-math></inline-formula>.
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.
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.000 | 0.001 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.001 | 0.002 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.000 |
| Open science | 0.004 | 0.002 |
| Research integrity | 0.001 | 0.001 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
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 it