Study on attacks against federated learning

• Main Author: Guo, Feiyan
• Other Authors: Yeo Chai Kiat
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2022
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: https://hdl.handle.net/10356/163119
• Institution: Nanyang Technological University

With the rise of artificial intelligence, the need for data also increases. However, many strict data privacy laws were put in place to protect personal data from being leaked. Therefore, this greatly limited the usage of artificial intelligence. Federated learning is a new form of collaborative machine learning that leverages on decentralized data for training models. This introduces the possibility of being exposed to poisoned data from malicious participants. In this project, the author explores different attack and defence methodologies to get a better understanding of how federated learning works. The focus is on the coordinated backdoor attack with model-dependant triggers for attack methodology and robust learning rates for defence methodology. The defence methodology is implemented into an opensourced federated learning base code. This will allow federated learning to be more widely used since it is less likely to be compromised by malicious attackers in the presence of built-in defences.