SketchGuard scales decentralized federated learning with sketch-based screening

Researchers have developed SketchGuard, a new framework designed to enhance the security and scalability of decentralized federated learning. This approach uses compressed sketches of model updates instead of full model vectors for initial screening, significantly reducing communication costs, especially when dealing with malicious participants. Experiments demonstrate that SketchGuard maintains state-of-the-art robustness against Byzantine attacks while achieving substantial computational savings. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Enhances security and efficiency for collaborative AI model training in decentralized environments.

RANK_REASON Academic paper detailing a new method for improving decentralized federated learning.

Read on arXiv cs.LG →

• SketchGuard
• Murtaza Rangwala
• arXiv
• Count Sketch

• paper
• safety