Data Reconstruction and Protection in Federated Learning for Fine-Tuning Large Language Models
Bibliographic record
Abstract
Federated learning can facilitate multiple parties to train a shared model on their own private data in a communication-efficient manner. It offers significant benefits for fine-tuning pre-trained large language models, as it supports distributed fine-tuning with a wider range of diverse data while preserving data privacy. However, recent research has revealed a potential privacy vulnerability in federated learning, specifically in the sharing of gradients from clients to server. This vulnerability can lead to the leakage of training data for Transformer-based large language models, thereby allowing the recovery of textual data. In this paper, we conduct a comprehensive evaluation of the effectiveness of the state-of-the-art gradient leakage attacks on textual data within the context of fine-tuning large language models. Our findings reveal that the key element for the attack's success — the target gradient — is not as readily obtainable for the adversary as previously assumed, particularly in regards to the Transformer architecture and practical federated learning settings. A technical error in their implementations has inadvertently caused the gradient to become more associated with the target data than intended. With the error fixed and when following the conventional federated learning framework, gradient leakage attacks pose minimal threats to large language models.
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 | 0.036 |
| Meta-epidemiology (narrow) | 0.001 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.001 | 0.004 |
| Scholarly communication | 0.003 | 0.007 |
| Open science | 0.002 | 0.008 |
| Research integrity | 0.002 | 0.005 |
| 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".