Outsourcing Training without Uploading Data via Efficient Collaborative Open-Source Sampling.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Cloud Computing

Background:

• Deep learning demands significant computational and data resources, making cloud-based training attractive.
• Uploading sensitive client data to cloud servers for training poses privacy and bandwidth challenges.

Purpose of the Study:

• To propose a novel strategy for outsourcing deep learning model training to the cloud without uploading sensitive client data.
• To leverage open-source data as a proxy for client data in cloud-based training.

Main Methods:

• Developed Efficient Collaborative Open-source Sampling (ECOS) to construct a proxy dataset from open-source data.
• ECOS probes open-source data to infer client data distribution through efficient sampling.
• The process involves communicating compressed public features and client scalar responses.

Main Results:

• ECOS effectively constructs a proximal proxy dataset from open-source data for cloud training.
• Demonstrated improvements in automated client labeling, model compression, and label outsourcing.
• Empirical studies confirmed ECOS's effectiveness across various learning scenarios.

Conclusions:

• ECOS offers a viable solution for privacy-preserving and efficient cloud-based deep learning model training.
• Leveraging open-source data mitigates the need for sensitive client data transfer.
• The proposed method enhances collaborative learning by enabling secure outsourcing.