| 研究課題/領域番号 |
21J23090
|
|---|
| 配分区分 | 補助金 |
|---|
| 研究機関 | 京都大学 |
|---|
研究代表者 |
鄭 舒元 京都大学, 情報学研究科, 特別研究員(DC1)
|
|---|
| 研究期間 (年度) |
2021-04-28 – 2024-03-31
|
| キーワード | data market / privacy protection / data valuation |
|---|
| 研究実績の概要 |
We have made a significant advance in building a reliable data market for machine learning (ML) applications. Concretely, in a collaborative data marketplace where multiple data owners collaborate to train an ML model, it is essential to evaluate the owners' diverse contributions to the model's utility to encourage fair cooperation. However, existing studies have neglected the potential privacy leakage in the contribution evaluation process. We have proposed pioneering methods for privacy-preserving contribution evaluation in collaborative ML to address this significant limitation. Our methods enable buyers to estimate data products' qualities before purchasing without accessing the products and sacrificing data owners' privacy, which considerably promotes reliable data trading.
|
| 現在までの達成度 (区分) |
現在までの達成度 (区分)
2: おおむね順調に進展している
理由
We have completed this fiscal year's research plan as scheduled with high quality. We designed a fair, efficient, and privacy-preserving method for evaluating contributions in data trading, which we have published in the top-tier journal PVLDB. Furthermore, we presented the work accomplished in the previous year at the prestigious international conference IEEE BigData 2022.
|
| 今後の研究の推進方策 |
We will initiate a research task on how to prevent data buyers from stealing ML models from prediction API marketplaces. Firstly, we need to investigate and characterize potential adversaries who may attempt model stealing attacks in the prediction API market and define our goals to formulate our research problem. Then, we plan to tackle this challenging problem using a game-theoretical approach. Solving this problem will have significant positive implications for building a reliable and sustainable data marketplace.
|
