Federated learning vulnerable to privacy leaks and data poisoning attacks

A new report details the vulnerabilities of federated learning, a method for adapting pretrained models to local user data. Researchers found that a central server can recover private label sets from a single client update, though reconstructions may lack detail. While label-count recovery decreases with more clients, backdoor attacks remain a significant threat, with success rates increasing as more clients become malicious. Defense mechanisms like FLTrust can mitigate these risks, but are ineffective when all clients are compromised. AI

IMPACT Highlights critical security and privacy risks in federated learning, necessitating robust defense mechanisms for widespread adoption.

RANK_REASON Research paper detailing vulnerabilities in federated learning. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• FedAvg
• Federated Art Chase
• federated learning
• FedMaestro
• FLTrust
• Google Colab
• ImageNet
• Indigma
• MobileNetV4-Conv-Small
• Quick, Draw!
• Vasileios Perifanis

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →