Towards Understanding Quality Challenges of the Federated Learning: A First Look from the Lens of Robustness

Amin Eslami Abyane, Derui Zhu, Roberto Medeiros de Souza, Lei Ma, Hadi Hemmati

Introduction

Federated Learning (FL) has emerged as a cornerstone for privacy-preserving machine learning, enabling model training on decentralized data without requiring raw data transfer. However, this decentralization introduces significant security risks. Because the central server cannot inspect local training processes or raw data, FL is inherently vulnerable to Byzantine faults—a category encompassing both unintentional system failures and malicious adversarial attacks.

The Federated Learning Framework

In a typical FL setup, a central server (the aggregator) coordinates training across multiple clients. Each client trains a local model on its private data and sends the resulting updates back to the server. The server then aggregates these updates to produce a global model.

Overview of the Federated Learning architecture with a central aggregator and decentralized local models.

While standard Federated Averaging (FedAvg) is efficient, it lacks robustness against malicious updates. This study evaluates several "robust" alternatives, including Krum (distance-based), Median (coordinate-wise), and Trimmed Mean (outlier removal).

Methodology: The Ensemble Approach

The researchers conducted a massive empirical study involving 4,960 models across 496 configurations. A key contribution is the introduction of an Ensemble Aggregator. Instead of committing to a single defense, the server runs multiple aggregators in parallel and selects the update that yields the highest validation accuracy on a small, clean proxy dataset held by the server.

Threat Model and Data Faults

The study categorizes quality challenges into two main groups:

  1. Adversarial Attacks: These include data poisoning (e.g., label flipping), model poisoning (e.g., sign flipping), and targeted backdoor attacks.
  2. Simulated Faults (Mutations): These represent real-world data issues like Gaussian noise, class overlap, and data imbalance.
Example of a backdoor attack trigger (a white 'Z' pattern) applied to an image to misclassify it.

Backdoor attacks are particularly insidious because they embed a specific trigger (like the 'Z' pattern shown above) that causes the model to misclassify specific inputs while maintaining high accuracy on the main task.

Key Empirical Insights

  • Model Poisoning vs. Data Poisoning: Directly manipulating model gradients is far more destructive than simply flipping labels. Model poisoning can instantly degrade a global model to the level of random guessing.
  • The Non-IID Challenge: In heterogeneous (Non-IID) data environments, robust aggregators like Krum often struggle. They frequently misidentify legitimate data diversity as an attack, leading to significant performance drops even in benign settings.
  • Inherent Resilience: Surprisingly, FL is naturally resilient to common data noise (e.g., camera artifacts). The aggregation process itself acts as a filter, with most non-malicious faults impacting accuracy by less than 6%.

Results and Performance

The Ensemble Aggregator proved to be the most versatile defense, outperforming individual aggregators in 75% of the tested scenarios. It successfully "switches" its defense strategy based on the observed threat.

However, specific aggregators still hold niche advantages. For instance, Krum was the only method capable of effectively suppressing backdoor attacks, keeping backdoor success rates near zero while other methods allowed them to reach nearly 100%. On the other hand, FedAvg remains superior for clean, well-distributed data but is highly vulnerable to model poisoning, where its accuracy can drop from 77% to 10%.

Conclusion and Future Work

While the Ensemble Aggregator offers a robust general-purpose defense, it introduces a ~50% computational overhead. Furthermore, it currently struggles with backdoors that do not degrade main-task accuracy. Future research will focus on more efficient ensemble heuristics and developing "attacker-aware" defenses that can anticipate adversarial adaptations.

Summary updated

Abstract

Federated learning (FL) is a distributed learning paradigm that preserves users’ data privacy while leveraging the entire dataset of all participants. In FL, multiple models are trained independently on the clients and aggregated centrally to update a global model in an iterative process. Although this approach is excellent at preserving privacy, FL still suffers from quality issues such as attacks or byzantine faults. Recent attempts have been made to address such quality challenges on the robust aggregation techniques for FL. However, the effectiveness of state-of-the-art (SOTA) robust FL techniques is still unclear and lacks a comprehensive study. Therefore, to better understand the current quality status and challenges of these SOTA FL techniques in the presence of attacks and faults, we perform a large-scale empirical study to investigate the SOTA FL’s quality from multiple angles of attacks, simulated faults (via mutation operators), and aggregation (defense) methods. In particular, we study FL’s performance on the image classification tasks and use Deep Neural Networks as our model type. Furthermore, we perform our study on two generic image datasets and one real-world federated medical image dataset. We also systematically investigate the effect of the proportion of affected clients and the dataset distribution factors on the robustness of FL. After a large-scale analysis with 496 configurations, we find that most mutators on each user have a negligible effect on the final model in the generic datasets, and only one of them is effective in the medical dataset. Furthermore, we show that model poisoning attacks are more effective than data poisoning attacks. Moreover, choosing the most robust FL aggregator depends on the attacks and datasets. Finally, we illustrate that a simple ensemble of aggregators achieves a more robust solution than any single aggregator and is the best choice in 75% of the cases. The data that support the findings of this study are available in our repository at https://github.com/aminesi/federated .
BibTeX
@inproceedings{abyane2023towards,
  author = {Amin Eslami Abyane and Derui Zhu and Roberto Medeiros de Souza and Lei Ma and Hadi Hemmati},
  title = {{Towards Understanding Quality Challenges of the Federated Learning: A First Look from the Lens of Robustness}},
  booktitle = {Empir. Softw. Eng},
  year = {2023},
  url = {https://openreview.net/forum?id=NQ3kRQirVw},
}
Figure