Core Concepts
Decoupled Vertical Federated Learning (DVFL) offers fault tolerance and privacy in training, outperforming traditional methods.
Abstract
Abstract:
Vertical Federated Learning (VFL) overview.
Introduction to DVFL as a fault-tolerant approach.
Decoupled VFL Approach:
Training process and fault tolerance.
Isolation between feature learning and label supervision.
Data Extraction:
"DVFL allows for decentralized aggregation and isolation between feature learning and label supervision."
"Model performance is comparable to VFL on various classification datasets."
Experiments:
Fault simulation and performance under different scenarios.
Impact of redundancy on model performance.
Training with limited intersection data.
Security & Privacy:
Elimination of gradient-based inference attacks.
Conclusion & Future Work:
Summary of DVFL benefits and potential future directions.
Stats
"DVFL allows for decentralized aggregation and isolation between feature learning and label supervision."
"Model performance is comparable to VFL on various classification datasets."
Quotes
"DVFL allows for decentralized aggregation and isolation between feature learning and label supervision."
"Model performance is comparable to VFL on various classification datasets."