Abstract: Federated learning (FL) facilitates collaborative model training across distributed clients while maintaining privacy. Federated noisy label learning (FNLL) is more of a challenge for data inaccessibility and noise heterogeneity. Existing works primarily assume clients are either noisy or clean, which may lack the flexibility to adapt to diverse label noise across different clients, especially when entirely clean or noisy clients are not the majority. To address this, we propose a general noise-robust federated learning framework called Federated Early-Stopping (FedES), which adaptively updates critical parameters of each local model based on their noise rates, thereby avoiding overfitting to noisy labels. FedES is composed of two stages: federated noise estimation and parameter-adaptive local updating & global aggregation. We introduce a signed distance based on local and global gradients during a federated round to estimate clients' noise rates without requiring additional information. Based on this measure, we employ various degrees of early-stopping during local updating on the clients, and further, a noise-aware global aggregation is employed to achieve noise-robust learning. Extensive experiments conducted on varying synthetic and real-world label noise demonstrate the superior performance of FedES over the state-of-the-art methods.
Introduction
Federated Learning (FL) allows multiple parties to train a shared model without exchanging raw data. However, real-world datasets are often "noisy" (mislabeled). In a federated setting, this noise is heterogeneous: one participant might have highly accurate labels, while another might have a high error rate.
Existing solutions for Federated Noisy Label Learning (FNLL) often treat clients as either entirely clean or entirely noisy. This binary approach frequently discards noisy clients altogether, leading to a loss of valuable clean data hidden within those sets. Furthermore, deep neural networks (DNNs) exhibit a "memorization effect," where they learn clean patterns first but eventually overfit to noisy labels. The challenge is to prevent the model from memorizing noise on a per-client basis without knowing their true noise rates or discarding their data.
Methodology
The proposed FedES framework addresses these challenges through a two-stage process that moves beyond binary classification of clients.
1. Federated Noise Estimation
The framework uses Criticality of Parameters (CoP)—a metric calculated as $|gradient \times weight|$—to identify which parameters are most influential during training. It introduces a Signed Earth Mover’s Distance (EMD) to compare the distribution of a client's local CoPs against the global model's CoPs. This allows the server to estimate a "data quality" score ($\rho_n$) for each client. A positive signed distance indicates higher quality (cleaner) data, while a negative distance suggests higher noise.
2. Parameter-Adaptive Local Updating
Instead of stopping training for the whole model, FedES uses Selective Gradient Decay (SeGD). Based on the estimated noise rate, each client identifies a subset of "non-critical" parameters. These parameters are "early-stopped" (their updates are suppressed) to prevent them from memorizing noise, while "critical" parameters continue to learn the underlying clean patterns.
3. Noise-Aware Global Aggregation (NaAgg)
During the server's aggregation phase, the model weights updates based on the estimated data quality ($\rho_n$), ensuring that cleaner clients have a stronger influence on the global model's direction.
Experimental Evaluation
The researchers evaluated FedES on major benchmarks including CIFAR-10, CIFAR-100 (with synthetic symmetric and asymmetric noise), and Clothing1M (a large-scale real-world dataset with naturally occurring label noise).
Key Results
FedES consistently outperformed state-of-the-art (SOTA) methods like S-FedAvg and FedNoRo:
- Superior Accuracy: On CIFAR-10 (Non-IID with 50% noise), FedES achieved 68.11% accuracy, significantly higher than the baseline FedAvg (32.56%).
- Real-World Performance: On the Clothing1M dataset, FedES reached 73.03% accuracy, outperforming all other general and binary de-noising methods.
- Estimation Accuracy: As shown in the quality correlation plots, the "Signed EMD" approach provides a much more accurate estimate of data quality compared to simpler mean-based methods, tracking the true quality closely.
Discussion and Conclusion
FedES provides a "surgical" approach to early stopping by treating it as a regularization problem. By freezing non-critical weights for noisy clients, it effectively prevents the memorization of garbage data without requiring a clean benchmark dataset on the server.
While FedES requires a few initial "warm-up" rounds to stabilize parameter criticality estimates and adds some computational complexity to the aggregation step, its robustness to heterogeneity makes it highly effective for real-world FL. Future work could involve exploring how FedES handles feature noise in addition to label noise and further addressing the privacy implications of sharing detailed parameter importance metrics.
Summary updated