Privacy-enhanced knowledge transfer with collaborative split learning over teacher ensembles
Knowledge Transfer has received much attention for its ability to transfer knowledge, rather than data, from one application task to another. In order to comply with the stringent data privacy regulations, privacy-preserving knowledge transfer is highly desirable. The Private Aggregation of Teacher...
Saved in:
| Main Authors: | , , , , , |
|---|---|
| 其他作者: | |
| 格式: | Conference or Workshop Item |
| 語言: | English |
| 出版: |
2023
|
| 主題: | |
| 在線閱讀: | https://hdl.handle.net/10356/172524 |
| 標簽: |
添加標簽
沒有標簽, 成為第一個標記此記錄!
|
| 總結: | Knowledge Transfer has received much attention for its ability to transfer knowledge, rather than data, from one application task to another. In order to comply with the stringent data privacy regulations, privacy-preserving knowledge transfer is highly desirable. The Private Aggregation of Teacher Ensembles (PATE) scheme is one promising approach to address this privacy concern while supporting knowledge transfer from an ensemble of "teacher"models to a "student"model under the coordination of an aggregator. To further protect the data privacy of the student node, the privacy-enhanced version of PATE makes use of cryptographic techniques at the expense of heavy computation overheads at the teacher nodes. However, this inevitably hinders the adoption of knowledge transfer due to the highly disparate computational capability of teachers. Besides, in real-life systems, participating teachers may drop out of the system at any time, which causes new security risks for adopted cryptographic building blocks. Thus, it is desirable to devise privacy-enhanced knowledge transfer that can run on teacher nodes with relatively fewer computational resources and can preserve privacy with dropped teacher nodes. In this connection, we propose a dropout-resilient and privacy-enhanced knowledge transfer scheme, Collaborative Split learning over Teacher Ensembles (CSTE), that supports the participating teacher nodes to train and infer their local models using split learning. CSTE not only allows the compute-intensive processing to be performed at a split learning server, but also protects the data privacy of teacher nodes from collusion between the student node and aggregator. Experimental results showed that CSTE achieves significant efficiency improvement from existing schemes. |
|---|