Research question
Can a single malicious participant backdoor a federated model?
Key findings
- Model-replacement attacks let one participant implant backdoors in a round - Secure aggregation prevents detection of the malicious update
How To Backdoor Federated Learning
Demonstrated that federated learning's secure aggregation makes it easy for one client to implant a backdoor, a foundational attack on decentralised training.
Contribute
Spot an error, or know a record we should add? Members can propose additions and submit corrections.