VFL-Cafe: Communication-Efficient Vertical Federated Learning via Dynamic Caching and Feature Selection.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Vertical Federated Learning (VFL) enables privacy-preserving collaborative model training.
• High communication costs are a major challenge in VFL due to intermediate result transmission.
• Existing methods using stale results can reduce accuracy, especially with noisy data.

Purpose of the Study:

• To propose VFL-Cafe, a novel VFL training method.
• To enhance communication efficiency and model accuracy in VFL.
• To address limitations of current communication-efficient VFL approaches.

Main Methods:

• Dynamic caching of intermediate results for strategic reuse.
• Feature selection integrated into local updates to mitigate noisy features.
• Theoretical analysis for cache configuration optimization.

Main Results:

• VFL-Cafe significantly reduces communication overhead.
• The method maintains or improves model accuracy.
• Experimental results validate the efficacy of VFL-Cafe.

Conclusions:

• VFL-Cafe offers an effective solution for communication-efficient VFL.
• Dynamic caching and feature selection are key to improved performance.
• The proposed method balances efficiency and accuracy in VFL training.