Federated learning data corruption detection methods compared

A new research paper explores methods for detecting data corruption in federated learning environments. The study compares two signals: input-conditional uncertainty and prediction-label loss, evaluating their effectiveness against image noise and label flips. Results indicate that prediction-label loss is more effective for detecting persistent label flips, while uncertainty measures are better suited for identifying image noise. The paper suggests that robust federated learning data quality assessment should align the detection signal with the specific type of corruption present. AI

IMPACT This research could lead to more robust federated learning systems by improving data quality assessment and mitigating the impact of corrupted data.

RANK_REASON Research paper published on arXiv detailing methods for detecting data corruption in federated learning. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• CIFAR-10
• federated learning
• ResNet-20
• The Street View House Numbers Dataset

• paper
• safety

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