Robust Collaborative Learning with Data Silos

The School of EECS is hosting the following HDR Thesis Review 3 Seminar:

Robust Collaborative Learning with Data Silos

Speaker: Achmad Ginanjar

Host: Prof Xue Li

Chair: Dr Yadan Luo

Abstract: The expectation of government agencies to collaborate has become increasingly common with the advent of cloud technology. Cloud technology has effectively removed the barriers that previously hindered information sharing. However, Wyld and Maurin predicted that adopting cloud technology within government sectors would bring about coordination challenges. The presence of political correctness further complicates the sharing of information. Another study has also noted that the perception of data ownership among governmental agencies acts as a deterrent to information sharing.

When discussing collaboration within a constrained area such as health, government, or finance, data silos are a common scenario. Due to its scope and how it works, areas like the government hold massive amounts of data that are spread across several data silos. Collaborative learning in such environments faces three main problems, which are Learning, Privacy and Sharing (LPS) . The learning problems commonly faced with the massive size of data are very typical for government as data authorities. The second problem is privacy, which is always the centre of the issue with respect to collaboration and information sharing. The last problem is the sharing itself. How to make a sharing schema that every party agrees on in the sharing ecosystem.

We propose Robust Contrastive Federated Learning (RCFL). This final approach consists of four major concepts as a unified solution. The four concepts are Federated Learning, Tabular Contrastive Learning, Continual Learning and Random Selection. To combine these methods, technical improvements and adaptations were studied and applied. Our RCFL process involves three steps: local contrastive learning, global federated learning, and local continual learning.

Bio: Achmad Ginanjar received a Master of Data Science degree from Monash University, Australia in 2019. He is currently working toward a Ph.D. degree with the Electrical Engineering Department at the University of Queensland (UQ), Brisbane, QLD, Australia.

Before joining UQ, he is an IT professional from The Indonesia Tax Office and a data analytic lecture at Indonesia State College of Accountancy. His current research interests include machine learning, optimalisation, parallel algorithm.