Authors:

Avi Amalanshu

Date Created:

Not specified

Course Title:

Professor:

Not specified

About Paper:

Vertical Federated Learning (VFL) is a distributed machine learning paradigm wherein owners of disjoint features of a common set of entities collaborate to learn a global model while maintaining data privacy. This allows agents who individually have incomplete information about their target to learn meaningful joint representations when regulations or ethics disallow data sharing. For instance, an imaging center, pathology lab and OPD may come together to predict rare types of cancer, even though HIPAA norms may disallow them from collecting each others' data. In VFL, the "host" client owns data labels for each entity. The host learns a final representation based on intermediate local representations from all participants. Therefore, the host is a single point of failure for faults and attacks. Furthermore, the label feedback can be used by malicious "guest" clients for various inference attacks. Requiring the label owner to remain active and trustworthy during the entire training process is impractical and limits the applicability of VFL. We propose Decentralized VFL (DVFL), a blockwise approach to VFL that addresses these issues by decentralizing aggregation and decoupling guest training, aggregation and label supervision. We also introduce an asynchronous variant, Async-DVFL, which relaxes the entity alignment phase, viz. participants need not agree on the specific data sample they will process during a training round. This leads to a higher degree of privacy during training and makes implementation more practical. We show that the added redundancy due to decentralization greatly improves robustness and fault tolerance, and that decoupling the training process in a self-supervised manner does not sacrifice performance. We present experiments using this technique on standard machine learning tasks.

Source:

Purdue University / 2023

Topics:

No topics listed

Co-authors:

Avi Amalanshu