Different Corruptions, Different Signals: Uncertainty and Loss in Federated Data Quality
Loss caught bad labels; uncertainty mostly did not.
The paper tests federated-learning corruption signals on CIFAR-10 and SVHN with non-IID client partitions. For persistent random label flips, prediction-label loss reached per-sample AUCs of 0.85 and 0.95, while uncertainty estimates stayed near chance. For additive image noise, expected-entropy uncertainty rose above chance and slightly beat loss. The authors argue FL data-quality checks should choose the signal based on the corruption type, rather than treating uncertainty as a general detector. ArXiv · AI/CL/LG's note
The paper tests federated-learning corruption signals on CIFAR-10 and SVHN with non-IID client partitions. For persistent random label flips, prediction-label loss reached per-sample AUCs of 0.85 and 0.95, while uncertainty estimates stayed near chance. For additive image noise, expected-entropy uncertainty rose above chance and slightly beat loss. The authors argue FL data-quality checks should choose the signal based on the corruption type, rather than treating uncertainty as a general detector. ArXiv · AI/CL/LG's note
score 3