FLOSS: Federated Learning with Opt-Out and Straggler Support
A new system called FLOSS addresses data privacy challenges in federated learning by handling user opt-out and straggler devices. While prior work focused on privacy-preserving operations for consenting users, modern agreements allow users to opt out of sharing data while still using the system. Combined with stragglers from heterogeneous device capabilities, this leads to missing data that introduces bias and degrades model performance. FLOSS mitigates these impacts, as demonstrated in simulations.
Key facts
- FLOSS stands for Federated Learning with Opt-Out and Straggler Support
- Previous work on data privacy in federated learning focused on privacy-preserving operations for users who agreed to share data
- Modern data privacy agreements empower users to opt out of sharing data while using the system
- Stragglers arise from heterogeneous device capabilities
- Missing data from opt-out and stragglers introduces bias and degrades model performance
- FLOSS mitigates the impacts of missing data on federated learning
- The system's performance was empirically demonstrated in simulations
- The paper is categorized under Computer Science > Machine Learning
Entities
Institutions
- arXiv