Abstract
Federated Adversarial Learning (FAL) is a robust framework for resisting adversarial attacks on federated learning. Although some FAL studies have developed efficient algorithms, they primarily focus on convergence performance and overlook generalization. Generalization is crucial for evaluating algorithm performance on unseen data. However, generalization analysis is more challenging due to non-smooth adversarial loss functions. A common approach to addressing this issue is to leverage smoothness approximation. In this paper, we develop algorithm stability measures to evaluate the generalization performance of two popular FAL algorithms: Vanilla FAL (VFAL) and Slack FAL (SFAL), using three different smooth approximation methods: 1) Surrogate Smoothness Approximation (SSA), (2) Randomized Smoothness Approximation (RSA), and (3) Over-Parameterized Smoothness Approximation (OPSA). Based on our in-depth analysis, we answer the question of how to properly set the smoothness approximation method to mitigate generalization error in FAL. Moreover, we identify RSA as the most effective method for reducing generalization error. In highly data-heterogeneous scenarios, we also recommend employing SFAL to mitigate the deterioration of generalization performance caused by heterogeneity. Based on our theoretical results, we provide insights to help develop more efficient FAL algorithms, such as designing new metrics and dynamic aggregation rules to mitigate heterogeneity.
Introduction
Federated Learning (FL) enables multiple parties to collaboratively train a global model without sharing private local data. Despite its benefits, FL models are often vulnerable to adversarial attacks—small, intentional perturbations to input data that can cause catastrophic model failure. Federated Adversarial Learning (FAL) has emerged as a key defense mechanism. However, while existing research has focused on the convergence of FAL algorithms, their generalization performance—how well they perform on unseen data—remains poorly understood. Analyzing generalization in FAL is mathematically difficult because adversarial loss functions are typically non-smooth, breaking standard theoretical tools.
Methodology: Smoothness Approximation in FAL
To address the non-smoothness of adversarial loss, this study evaluates three primary smoothness approximation methods to determine their impact on the generalization gap:
- Surrogate Smoothness Approximation (SSA): Replaces the non-smooth loss with a smoother surrogate function.
- Randomized Smoothness Approximation (RSA): Introduces random noise to model parameters during training to "blur" the loss landscape, creating a flatter and less sensitive surface.
- Over-Parameterized Smoothness Approximation (OPSA): Leverages the theory that highly over-parameterized (wide) neural networks naturally possess smoother loss landscapes.
These methods are applied to two FAL frameworks: Vanilla FAL (VFAL), which uses standard model averaging, and Slack FAL (SFAL), which employs dynamic weighting during aggregation to prioritize clients with better performance.
Key Theoretical and Empirical Insights
The Superiority of RSA
Among the tested smoothing methods, RSA provides the tightest generalization bounds. By fostering a flatter weight space, RSA reduces the model's sensitivity to specific training data quirks. Theoretical analysis shows that the generalization error for RSA scales at $O(T^{1/4})$, a significant improvement over SSA's $O(T \log T)$ scaling, where $T$ represents training rounds.
The "Curse of Width" in OPSA
While over-parameterization (OPSA) can aid optimization, it may paradoxically harm generalization in FAL. Excessive network width can exacerbate differences between clients, leading the global model to become biased toward specific local datasets rather than generalizing across the entire distribution.
Impact of Heterogeneity and Client Scaling
Data heterogeneity (non-IID data) is a primary driver of generalization error in FAL. As data becomes more diverse across clients, local adversarial examples become highly biased, widening the generalization gap. However, increasing the number of clients ($m$) helps mitigate this by diluting individual client biases.
SFAL as a Mitigating Strategy
SFAL's strategy of down-weighting "outlier" clients—those exhibiting high adversarial loss—effectively counters the negative effects of data heterogeneity. This results in a more stable and generalizable global model compared to standard VFAL.
Conclusion
This research identifies RSA as the most effective smoothing method for ensuring generalization in FAL. Furthermore, SFAL is recommended for handling highly heterogeneous data environments. While this study provides a robust theoretical foundation, future work should explore these dynamics in deeper neural networks and investigate new aggregation metrics, such as contrastive loss or adversarial penalties, to further bridge the generalization gap in real-world federated systems.
Summary updated