Abstract
Federated learning (FL) has emerged as a promising, massively distributed way to train a joint deep model over large amounts of edge devices while keeping private user data strictly on device. In this work, motivated from ensuring fairness among users and robustness against malicious adversaries, we formulate federated learning as multi-objective optimization and propose a new algorithm, FedMGDA+, that is guaranteed to converge to Pareto stationary solutions. FedMGDA+ is simple to implement, has fewer hyperparameters to tune, and refrains from sacrificing the performance of any participating user. We establish the convergence properties of FedMGDA+ and point out its connections to existing approaches. Extensive experiments on a variety of datasets confirm that FedMGDA+ compares favorably against state-of-the-art methods.
Introduction
Federated Learning (FL) allows for the training of powerful global models on private data residing on millions of edge devices, such as smartphones and sensors, without the data ever leaving the device. However, the standard optimization objective in FL—typically the arithmetic average of all users' losses (e.g., FedAvg)—suffers from a "tyranny of the majority." In heterogeneous (Non-IID) environments, a model that minimizes average loss may perform poorly for minority groups or users with unique data distributions. Furthermore, this averaging mechanism is vulnerable to malicious actors; a single attacker can "shout louder" than others by inflating their loss or gradient, effectively hijacking the global model.
Reframing FL as Multi-Objective Optimization
To address these issues of fairness and robustness simultaneously, we propose a shift in perspective: treating federated learning as a Multi-Objective Optimization (MOO) problem. Instead of a single global goal, the server attempts to satisfy $m$ different objectives—one for each participating user—simultaneously.
The FedMGDA+ Algorithm
The core of our approach is the Multiple Gradient Descent Algorithm (MGDA), adapted for the federated setting with several key innovations:
- Pareto Stationarity: Rather than seeking a single "best" average model, FedMGDA+ seeks a Pareto stationary solution. This is a point where no user's performance can be improved without degrading the performance of at least one other user, ensuring that no participant's local performance is sacrificed for the sake of the average.
- The $\epsilon$-Constraint: We introduce a tunable parameter ($\epsilon$) that allows the server to strike a balance between maximizing average performance (similar to FedAvg) and enforcing strict fairness (as in pure MGDA).
- Gradient Normalization: To prevent malicious users from dominating updates, FedMGDA+ normalizes all user updates to unit length. This prevents any single user from exerting disproportionate influence through gradient scaling.
- Proximal Interpretation: We provide a novel theoretical link between FL and "proximal averages," demonstrating that FedAvg and its variants are approximations of a specific mathematical structure that FedMGDA+ optimizes more directly.
Robustness and Fairness: A Unified View
A key insight of FedMGDA+ is that fairness and robustness are intrinsically linked through the MOO framework.
Robustness as a Byproduct: Because the algorithm seeks a common descent direction that benefits the collective, it naturally ignores "outlier" updates from attackers that do not align with the honest majority. This makes robustness essentially "free" within the fairness framework.
Immunity to Inflation: Standard fairness methods are often vulnerable to "additive inflation," where a user adds a constant to their loss to gain more weight in the optimization. FedMGDA+ is immune to this because it centers the optimization around the previous iteration's values, making it invariant to such manipulations.
Experimental Evaluation
We validated FedMGDA+ across a diverse set of benchmarks, including CIFAR-10, Fashion-MNIST, Federated EMNIST (FEMNIST), the Shakespeare language modeling dataset, and the Adult census dataset.
Key Results
- Superior Fairness: On the FEMNIST dataset (3,400+ users), FedMGDA+ achieved a higher average accuracy (87.6%) compared to FedAvg (84.9%), while significantly reducing the variance across users. This indicates more consistent performance for all participants.
- Enhanced Robustness: In experiments involving "scaling attacks," where an adversary attempts to hijack the model, FedMGDA+ remained stable. In contrast, the performance of other state-of-the-art methods like AFL and $q$-FedAvg collapsed.
- User Improvement Rate: FedMGDA+ consistently showed a higher "improvement rate"—the percentage of users whose local model improved after a global update—compared to FedAvg, particularly as training progressed.
Conclusion and Future Work
FedMGDA+ offers a principled, simple-to-implement solution for fair and robust federated learning. By ensuring Pareto stationarity, it provides a stronger incentive for users to participate in the federated process. While the current implementation is efficient for hundreds of users, future work will explore more efficient approximation methods (such as Frank-Wolfe) for settings with millions of concurrent participants, as well as formal quantifications of privacy preservation when using normalized gradients.
Paper Citation: Zeou Hu, Kiarash Shaloudegi, Guojun Zhang, and Yaoliang Yu. "Federated Learning Meets Multi-objective Optimization." arXiv:2006.11489v2 [cs.LG], Jan 2023.
Summary updated