Robust Combination of Distributed Gradients Under Adversarial Perturbations

Kwang In KimOpenReview 2022 · OpenReview

Abstract

We consider distributed (gradient descent-based) learning scenarios where the server combines the gradients of learning objectives gathered from local clients. As individual data collection and learning environments can vary, some clients could transfer erroneous gradients eg, due to adversarial data or gradient perturbations. Further, for data privacy and security, the identities of such affected clients are often unknown to the server. In such cases, naively aggregating the resulting gradients can mislead the learning process. We propose a new server-side learning algorithm that robustly combines gradients. Our algorithm embeds the local gradients into the manifold of normalized gradients and refines their combinations via simulating a diffusion process therein. The resulting algorithm is instantiated as a computationally simple and efficient weighted gradient averaging algorithm. In the experiments with five classification and three regression benchmark datasets, our algorithm demonstrated significant performance improvements over existing robust gradient combination algorithms as well as the baseline uniform gradient averaging algorithm.

Introduction

In distributed learning environments, such as Federated Learning, a central server coordinates training across numerous local clients. The server distributes a global model, clients compute gradients based on their local datasets, and the server aggregates these updates. However, this architecture is highly vulnerable to "poisoned" gradients originating from malicious actors or clients with noisy data. Standard aggregation methods like FedAvg (uniform averaging) are notoriously sensitive to these outliers, which can significantly degrade model performance. While existing robust alternatives exist, they often require restrictive conditions, such as the server possessing a clean validation dataset or having prior knowledge of the exact number of compromised clients.

Methods: Diffusion on a Manifold

The proposed algorithm addresses these vulnerabilities by treating the collection of local gradients as a noisy sample from a manifold. The approach consists of several key steps:

  1. Manifold Embedding: Local gradients are normalized and embedded into a manifold of normalized gradients.
  2. Diffusion as Denoising: The initial uniform average of gradients is treated as a "noisy" observation. The algorithm then simulates a diffusion process using a discrete version of the Laplace-Beltrami operator to refine the combination weights.
  3. Iterative Weighting: Instead of simple averaging, the server calculates a weighted average. As optimization proceeds, the weights for "genuine" gradients (which form a consensus) are maintained, while the weights for "perturbed" or outlier gradients are automatically suppressed.
  4. Efficiency: Despite its mathematical complexity, the resulting algorithm is computationally efficient, scaling linearly ($O(K \times m)$) with the number of clients $K$ and the gradient vector size $m$.

Experimental Results

The algorithm was evaluated across five classification tasks (e.g., CIFAR10, FashionMNIST) and three regression benchmarks (e.g., MOCAP). The researchers simulated adversarial scenarios including random labeling and Gaussian gradient perturbations.

Comparison of FedAvg and the proposed diffusion-based aggregation on a synthetic classification task. The bottom plot illustrates the algorithm's ability to identify and down-weight adversarial clients (indices 80-120).

As shown in the figure above, the proposed method significantly outperforms FedAvg in recovering the underlying data distribution. Quantitatively, on CIFAR10 with 60% corrupted clients, the proposed method maintained 88.85% accuracy, whereas FedAvg dropped to 84.83% and Median-based Thresholding (MedTh) collapsed to 31.40%. The method also demonstrated "graceful degradation," remaining functional even when more than half of the clients were adversarial.

Conclusion

This research introduces a robust, privacy-preserving framework for gradient aggregation in distributed learning. By leveraging manifold geometry and diffusion processes, the algorithm effectively filters out adversarial noise without requiring external clean data or prior knowledge of attacker counts. Future work will explore handling partially corrupted clients and automating hyperparameter tuning for the diffusion step size and kernel width.

Summary updated

Abstract

We consider distributed (gradient descent-based) learning scenarios where the server combines the gradients of learning objectives gathered from local clients. As individual data collection and learning environments can vary, some clients could transfer erroneous gradients eg, due to adversarial data or gradient perturbations. Further, for data privacy and security, the identities of such affected clients are often unknown to the server. In such cases, naively aggregating the resulting gradients can mislead the learning process. We propose a new server-side learning algorithm that robustly combines gradients. Our algorithm embeds the local gradients into the manifold of normalized gradients and refines their combinations via simulating a diffusion process therein. The resulting algorithm is instantiated as a computationally simple and efficient weighted gradient averaging algorithm. In the experiments with five classification and three regression benchmark datasets, our algorithm demonstrated significant performance improvements over existing robust gradient combination algorithms as well as the baseline uniform gradient averaging algorithm.
BibTeX
@misc{kim2022robust,
  author = {Kwang In Kim},
  title = {{Robust Combination of Distributed Gradients Under Adversarial Perturbations}},
  year = {2022},
  url = {https://openreview.net/forum?id=wDr81QvEdgH},
  howpublished = {OpenReview},
}
Figure