[NextGen] Secure and Privacy-Enhancing Federated Learning: Algorithms, Framework, and Applications to NLP and Medical AI

Federated learning (FL) is an emerging machine learning paradium to enable distributed clients (e.g., mobile devices) to jointly train a machine learning model without pooling their raw data into a centralised server. Because data never leaves from user clients, FL systematically mitigates privacy risks from centralised machine learning and naturally comply with rigorous data privacy regulations, such as GDPR and Privacy Act 1988. 

FL has drawn significant attention recently from both research communities and industry sectors, and has been applied in real-world applications, e.g., keyboard suggestions [1] and emoji prediction [2]. Despite its great potential, FL faces critical security, privacy, and performance challenges under practical constraints and considerations. For example, poisioning attacks against FL can heavily impact model performance and cause model misclassification. In medical applications, such attacks might result in miss-diagnosis, lead to incorrect treatment and even death. Another noteworthy example is gradient leakage attack, where the orignial client data can be recovered via model gradients submitted from clients to the server. Such an attack directly breaks the claimed privacy guarantees of FL.

In this project, we aim at an ambitious goal - designing secure and privacy-enhancing algorithms and framework for FL and applying our designs into real-world applications. To achieve this goal, we expect to deeply and comprehensively explore the following directions:

1) Investigate privacy-enhancing techniques (e.g., differential privacy and/or secure multi-party computation) and design protocols to secure model updates throughout the process of FL collaborative training;

2) Investigate and design security measures (e.g., robust aggregation rules) to defeat model poisioning attacks against FL;

3) Design and optimise task-specific FL algorithms for NLP and Medical AI;

4) Design and implement a user-friendly FL framework for both research and technology adoption purposes.

We expect to recruit a cohort of students with diverse background and skill sets in machine learning, cybersecurity, and digital health to work in the above directions in a collaborative manner. This project will provide multiple competitive scholarships funded by a national elite HDR training program: Data61 Next Generation Graduate (https://www.csiro.au/en/work-with-us/funding-programs/programs/next-generation-graduates-programs),  and under a project titled as “Privacy-preserving Machine Learning: Technology Develeopment and Adoption” led by Monash Faculty of IT.  

Reference: 

[1] ”Federated Learning: Collaborative Machine Learning without Centralized Training Data”, https://ai.googleblog.com/2017/04/federated-learning-collaborative.html 

[2] “Learning with Privacy at Scale”, https://machinelearning.apple.com/research/learning-with-privacy-at-scale

[3] Fang et al., "Local Model Poisoning Attacks to Byzantine-Robust Federated Learning". In USENIX Security Symposium, 2020. 

[4] Zhu et al., "Deep Leakage from Gradients." Advances in Neural Information Processing Systems, 2019.