Is Split Learning Privacy-Preserving for Fine-Tuning Large Language Models?
Bibliographic record
Abstract
With the success of pre-trained large language models in various tasks, users, individuals and enterprises alike, may need to fine-tune these models with their own datasets. Split learning was proposed to divide the model and place a portion on each user's own device, and intermediate results in each iteration of training will be sent to the server to complete the forward pass. There were concerns in the literature about whether private data can be leaked by sending such intermediate results from the training process. In this paper, we conduct empirical studies on typical large language models, such as GPT-2, OPT, Llama, and Qwen, to show that in most situations, an honest-but-curious server is not able to reconstruct private data using such intermediate results. To find out the reason why large language models preserve data privacy better in these situations, we present our theoretical analyses on these empirical observations. In one special case, where a state-of-the-art existing attack can reconstruct data in the first iteration, we show that it can be easily defended with a simple but effective solution leveraging publicly accessible data.
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 | 0.016 | 0.066 |
| Meta-epidemiology (narrow) | 0.001 | 0.001 |
| Meta-epidemiology (broad) | 0.002 | 0.002 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.002 | 0.006 |
| Scholarly communication | 0.004 | 0.018 |
| Open science | 0.004 | 0.008 |
| Research integrity | 0.003 | 0.007 |
| Insufficient payload (model declined to judge) | 0.003 | 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".