Alibaba open sources privacy-preserving computation 

Alibaba Group’s global research initiative Alibaba DAMO Academy has made the source code of its latest federated learning platform FederatedScope, a comprehensive platform with easy-to-use packages, accessible to the open-source community.

With all forms of Machine Learning (ML) and Artificial Intelligence (AI) on the rise, gathering training data to build and advance AI models is increasingly under the spotlight as the process could pose potential privacy concerns. 

To address such a challenge, federated learning – a way of privacy-preserving computation –– has emerged. By coordinating the training of micro-tasks across different end devices, intermediate training results – rather than raw user data – is fed back to the cloud server to alleviate privacy concerns. Yet it still enables data analytics and machine learning tasks across end devices.

DAMO stands for Discovery, Adventure, Momentum and Outlook.

“By sharing our self-developed federated learning technologies with the open-source community, we hope to promote the research and industrial deployment of privacy-preserving computation in different sectors, such as healthcare and smart mobility that usually involves sensitive user data and requires strict privacy protection practices,” said Bolin Ding, research scientist at Alibaba DAMO Academy.

Ding: Privacy-preserving computation is his thing.

In addition, with a newly-implemented event-driven framework, FederatedScope provides flexible support and comprehensive tools including a rich collection of benchmark datasets, well-known model architectures, advanced federated learning algorithms, easy-to-use automatic tuning functionalities and friendly interfaces. 

These enable researchers and developers to quickly build and customise task-specific federated learning applications in areas including computer vision, natural language processing, speech recognition, graph learning and recommendation.

For privacy protection in particular, the platform also offers cutting-edge technologies including differential privacy and multi-party computation to meet different requirements of privacy protection.

“We believe privacy-preserving computation is an important and essential trend,” added Ding. “Training AI models without compromising privacy is critical and that’s why we have devoted a lot of resources to drive the research of federated learning. We hope that by sharing our source codes and technology platform, we can support global developers in the community and encourage more innovation in this emerging field.”

According to Gartner, 60% of large organisations are expected to use one or more privacy-enhancing computation techniques by 2025.