2021-229 - Secure Aggregation in Federated Learning

Description:
  • Federated learning
  • Healthcare data
  • Financial data

Abstract

USC researchers have developed LightSecAgg, a novel approach for secure aggregation that guarantees privacy and dropout-resiliency while significantly cutting the overhead for resiliency against dropped users. Utilizing a “one-shot aggregate-mask reconstruction of the active users via mask encoding/decoding” technique, the approach can also be applied to secure aggregation in an asynchronous FL setting. In experiments using diverse training models and datasets, LightSecAgg significantly minimizes total training time.

Benefits

  • Privacy and dropout-resiliency parallels state-of-the-art
  • Decreases overhead for resiliency
  • Allows for scalable implementation
  • Increases speed of concurrent receiving and sending of chunked masks

Market Application

Federated learning (FL) allows for distributed learning over a large number of users while maintaining privacy, making it a promising approach for sensitive data applications such as healthcare and finance. FL employs secure model aggregation, which must be resilient to user dropouts. State-of-the-art protocols rely on secret sharing of random-seeds for mask generations to reconstruct and cancel dropped user models. However, this approach's complexity increases as the number of dropouts grows, making a scalable solution a key market need.

Publications

So, Jinhyun, et al. "LightSecAgg: a lightweight and versatile design for secure aggregation in federated learning." Proceedings of Machine Learning and Systems 4 (2022): 694-720. 

Stage of Development

  • Tested with diverse datasets and training models
  • Optimization in progress
  • Available for licensing

Patent Information:

  • Title: Systems and Methods for Improved Secure Aggregation in Federated Learning
  • App Type: Nationalized PCT
  • Country: United States
  • Serial No.: 18/682,656
  • Patent No.:  
  • File Date: 2/9/2024
  • Issued Date:  
  • Expire Date:  
  • Patent Status: Patent Pending