Abstract
Personalized federated learning (PFL) is designed for scenarios with non-independent and identically distributed (non-IID) client data. Existing model mixup-based methods, one of the main approaches of PFL, can only extract either global or personalized features during training, thereby limiting effective knowledge sharing among clients. To address this limitation, we propose the Dual Calibration-based PFL (DC-PFL). It divides local models into a heterogeneous feature extractor and a homogeneous classifier. The FL server utilizes mean and covariance representations from clients' feature extractors to train a global generalized classifier, facilitating information exchange while preserving privacy. To enhance personalization and convergence, we design a feature extractor-level calibration method with an auxiliary loss for local models to refine feature extractors using global knowledge. Furthermore, DC-PFL refines the global classifier through the global classifier-level calibration, utilizing sample representations derived from an approximate Gaussian distribution model specific to each class. This method precludes the need to transmit original data representations, further enhancing privacy preservation. Extensive experiments on widely used benchmark datasets demonstrate that DC-PFL outperforms eight state-of-the-art methods, surpassing the best-performing baseline by 1.22% and 9.22% in terms of accuracy on datasets CIFAR-10 and CIFAR-100, respectively.
Introduction
Federated Learning (FL) has emerged as a powerful paradigm for training AI models on distributed private data. However, real-world data is often statistically heterogeneous (non-IID), where every user's data distribution differs significantly. This challenge necessitates Personalized Federated Learning (PFL), which aims to tailor models to individual client needs while still benefiting from collective knowledge.
Standard PFL methods often struggle to balance global generalization with local personalization. Many existing approaches either over-prioritize global features, losing local nuances, or focus too heavily on local features, which hinders effective collaborative learning. Additionally, system heterogeneity—where clients have different hardware capabilities—often forces a one-size-fits-all model architecture. DC-PFL addresses these gaps by decoupling the model architecture and introducing a novel dual calibration strategy.
The DC-PFL Framework
The core innovation of DC-PFL lies in its ability to allow clients to maintain unique, heterogeneous feature extractors while sharing a high-performing global classifier. This is achieved through a four-step dual calibration process.
- Local Model Training & Feature Calibration: Clients train local models using their private data. An auxiliary loss is introduced to align local feature representations with global class means provided by the server, ensuring local extractors benefit from global insights.
- Class Feature Representation: To preserve privacy, clients do not share raw data or model weights. Instead, they calculate and transmit the mean and covariance of features for each class in their local dataset.
- Global Classifier Training: The server aggregates these class-wise statistics to update a global classifier, capturing a comprehensive view of the data distribution across all clients.
- Global Classifier Calibration: The server uses the received statistics to generate "virtual representations" from an approximate Gaussian distribution. It then fine-tunes the global classifier on these virtual samples, effectively performing data augmentation in the latent space to correct for local biases without ever seeing raw images.
Experimental Setup
The effectiveness of DC-PFL was evaluated on two standard computer vision benchmarks modified to simulate extreme non-IID conditions:
- CIFAR-10: Each client was restricted to only 2 out of 10 possible classes.
- CIFAR-100: Each client was restricted to only 10 out of 100 possible classes.
The data for each client was split into 80% for training, 10% for evaluation, and 10% for testing. DC-PFL was compared against eight state-of-the-art baselines, including FedAvg, FedProto, and FedGH.
Results and Analysis
DC-PFL demonstrated significant performance gains over existing methods. It achieved a 1.22% accuracy improvement on CIFAR-10 and a substantial 9.22% boost on CIFAR-100 compared to the best-performing baselines.
As shown in the figures above, DC-PFL (labeled as KDC-PFL) consistently maintains superior accuracy even as client participation rates decrease. This robustness is critical for real-world FL deployments where not all devices are online simultaneously. Ablation studies further confirmed that both the auxiliary loss (local calibration) and virtual sample generation (global calibration) are essential components, as removing either led to a noticeable decline in performance.
Conclusion
DC-PFL introduces a robust framework for personalized federated learning that balances privacy, architectural flexibility, and high accuracy. By sharing class-wise statistics rather than raw data, it mitigates privacy risks while enabling powerful global calibration through virtual sample generation. While the current approach assumes a Gaussian distribution for latent features, future work may explore more complex distribution models to further enhance performance in highly nuanced feature spaces.
Reference: Tang, X., Yu, H., Tang, R., Ren, C., Li, A., & Li, X. (2024). Dual Calibration-based Personalised Federated Learning. Proceedings of the 33rd International Joint Conference on Artificial Intelligence (IJCAI-24).
Summary updated