New Temperature Scaling Attack targets model confidence in federated learning

Researchers have developed a new training-time attack called the Temperature Scaling Attack (TSA) that specifically targets the confidence calibration of models in federated learning systems. This attack degrades a model's ability to accurately represent its own uncertainty while maintaining high predictive accuracy. TSA works by injecting temperature scaling with learning rate-temperature coupling during local training, which shifts model confidence without significantly altering accuracy or optimization signals. The attack has demonstrated substantial increases in calibration errors on benchmarks like CIFAR-100 and has shown potential for severe failures in critical applications such as healthcare and autonomous driving. AI

IMPACT This attack highlights a critical vulnerability in federated learning systems, potentially impacting the reliability of AI in sensitive applications.

RANK_REASON Research paper detailing a novel attack method on federated learning models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• CIFAR-100
• federated learning
• Kichang Lee
• Temperature Scaling Attack

• paper
• safety

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