FedFixer: Mitigating Heterogeneous Label Noise in Federated Learning (2403.16561v1)

Abstract: Federated Learning (FL) heavily depends on label quality for its performance. However, the label distribution among individual clients is always both noisy and heterogeneous. The high loss incurred by client-specific samples in heterogeneous label noise poses challenges for distinguishing between client-specific and noisy label samples, impacting the effectiveness of existing label noise learning approaches. To tackle this issue, we propose FedFixer, where the personalized model is introduced to cooperate with the global model to effectively select clean client-specific samples. In the dual models, updating the personalized model solely at a local level can lead to overfitting on noisy data due to limited samples, consequently affecting both the local and global models' performance. To mitigate overfitting, we address this concern from two perspectives. Firstly, we employ a confidence regularizer to alleviate the impact of unconfident predictions caused by label noise. Secondly, a distance regularizer is implemented to constrain the disparity between the personalized and global models. We validate the effectiveness of FedFixer through extensive experiments on benchmark datasets. The results demonstrate that FedFixer can perform well in filtering noisy label samples on different clients, especially in highly heterogeneous label noise scenarios.

• The paper introduces a dual model structure that alternates between personalized and global models to effectively differentiate clean from noisy labels in federated learning.
• It leverages a confidence regularizer to reduce overfitting on uncertain predictions, ensuring robust learning from clean data samples.
• Experimental results on benchmark datasets under various noise and distribution settings show FedFixer outperforms state-of-the-art methods with up to a 10% accuracy improvement.

Mitigating Heterogeneous Label Noise in Federated Learning with FedFixer

Overview

Federated Learning (FL) has emerged as a promising decentralized machine learning paradigm, enabling multiple clients to collaboratively train a global model under the coordination of a central server without sharing their local data. However, the presence of label noise, especially when heterogeneous across clients, significantly impairs the performance of FL models. In this context, Xinyuan Ji, Zhaowei Zhu, et al. introduce FedFixer, a novel approach designed to address heterogeneous label noise in FL by leveraging a dual model structure at the client level. The proposed method, validated on benchmark datasets, demonstrates remarkable effectiveness in filtering noisy label samples across varying clients, especially in highly heterogeneous label noise scenarios.

Key Contributions

• Dual Model Structure: FedFixer introduces a dual model structure at each client, comprising a personalized model and a global model. This structure enables the effective differentiation between clean client-specific samples and noisy label samples by allowing the models to alternately update based on samples selected by the other. This mechanism significantly reduces the risk of mislabeling client-specific samples as noisy labels.
• Confidence Regularizer (CR): To combat the challenge of models overfitting to noisy data, the authors incorporate a confidence regularizer. This regularizer diminishes the impact of unconfident predictions, which are often a result of label noise, thus enabling the model to fit better to clean data.
• Distance Regularizer (DR): A distance regularizer is also introduced to moderate the divergence between the personalized and global models. This component is crucial for preventing the personalized model from drifting too far as a result of overfitting to local noisy data.

Experimental Validation

FedFixer was extensively evaluated on multiple benchmark datasets, including MNIST, CIFAR-10, and Clothing1M, under both IID and non-IID data distributions with various levels of label noise. The results affirm the superiority of FedFixer in managing heterogeneous label noise, where it consistently outperforms state-of-the-art methods in high noise scenarios. Particularly notable is its performance in highly heterogeneous label noise conditions, where FedFixer can achieve up to 10\% improvement in accuracy over existing approaches.

Implications and Future Directions

The development of FedFixer provides a robust solution to one of the significant challenges in FL, opening up new avenues for the deployment of FL in real-world applications where data quality cannot be uniformly guaranteed. The dual model structure, alongside the thoughtful integration of confidence and distance regularizers, sets a foundational approach for future research focused on improving FL model resilience against various types of noise and data heterogeneity. Future work could explore the potential of integrating label correction mechanisms within the FedFixer framework to further enhance its efficacy, especially in scenarios with extreme noise levels.

Conclusion

The introduction of FedFixer marks a significant advancement in addressing the issue of heterogeneous label noise in federated learning. By innovatively employing a dual model structure complemented by confidence and distance regularizers, FedFixer not only mitigates the detrimental effects of label noise but also preserves the ability to learn from client-specific data. This approach paves the way for more robust and reliable FL applications, even in environments plagued by data quality issues.