Self-Driven Entropy Aggregation for Byzantine-Robust Heterogeneous Federated Learning

Wenke Huang, Zekun Shi, Mang Ye, He Li, Bo Du

TL;DR

Self-Driven Entropy Aggregation (SDEA) is a novel defense mechanism for Federated Learning (FL) that protects against Byzantine (malicious) attacks in heterogeneous data environments. Unlike previous methods, it does not require labeled or domain-specific proxy data. Instead, it uses the prediction entropy of models on a random public dataset to distinguish honest participants from attackers.

The Vulnerability of Federated Learning

Federated Learning allows multiple parties to train a shared model without exposing their private data. However, this decentralized approach is highly vulnerable to Byzantine attacks, where malicious clients deliberately upload "poisoned" updates to degrade the global model's performance.

Existing defenses struggle with two main issues:

  1. Data Heterogeneity (Non-IID): When honest clients have vastly different data distributions, their updates naturally diverge. Traditional defenses often mistake this natural variance for a malicious attack.
  2. The Proxy Data Trap: Many robust aggregation methods require a "gold standard" proxy dataset that is both labeled and closely related to the clients' private data. In real-world scenarios, obtaining such datasets is often expensive or impossible.

Introducing SDEA: Methods and Architecture

To address these limitations, the authors introduce SDEA. The framework leverages a key observation: deep neural networks trained on real data tend to be "over-confident" (producing sharp, low-entropy predictions) even on unrelated data, whereas malicious models often produce "smoother," high-entropy outputs.

SDEA Architecture: Using public data to calculate Instance Sharpness and Class Diversity, followed by weight clustering.

The SDEA framework consists of three main components:

  • Instance Sharpness (IS): The server feeds a random public dataset (like Tiny-ImageNet) into the aggregated global model. It optimizes learnable aggregation weights to minimize the entropy of individual predictions. This naturally assigns higher weights to "confident" (likely benign) clients.
  • Class Diversity (CD): To prevent the model from becoming biased toward just a few clients (the "Matthew Effect"), SDEA maximizes the entropy of the batch-average prediction. This ensures the model maintains a diverse set of predictive preferences across the public data.
  • Cooperative Cluster (CC): SDEA uses unsupervised clustering (specifically the FINCH algorithm) on the aggregation weights to separate clients into "benign" and "malicious" groups. It then averages the weights within the benign group to ensure fair contribution and stability.

Key Insights

  • Malicious Smoothness: Malicious updates—whether random noise or carefully crafted to stay "close" to the mean—tend to disrupt the structured over-confidence of a well-trained network. This manifests as higher entropy on public data probes.
  • Heterogeneous Sharpness: In Non-IID settings, different honest clients will be "sharp" on different types of images. If a system only minimizes entropy, it might accidentally ignore many honest clients. The Class Diversity term forces the server to incorporate a broader range of honest updates.
  • Randomness is a Feature: A "related" dataset is not required to probe a model. A model trained to recognize cats and dogs has internal structures that react predictably to images of cars or buildings, whereas a poisoned model does not.
Motivation: Benign clients (1-7) show lower entropy on public data compared to malicious attackers (8-9).

Datasets and Key Results

The authors evaluated SDEA across several standard benchmarks using private client data (CIFAR-10, MNIST, Fashion-MNIST) and random public data (Tiny-ImageNet, Market1501, SVHN, SYN). Crucially, the public data did not need to share the same label space as the private data.

SDEA was tested against both data-based attacks (Label Flipping) and parameter-based attacks (Random Noise, "A Little is Enough," and distance-minimization attacks like MiMa/MiSu). It consistently outperformed state-of-the-art robust aggregators like Multi Krum, Bulyan, and RFA, especially in highly heterogeneous settings.

  • Accuracy Boost: On CIFAR-10 with a "Random Noise" attack, SDEA achieved 69.21% accuracy, while the best competitor (DnC) reached only 64.72%.
  • Robustness to Attack Scale: Even when 40% of the clients were malicious, SDEA maintained high performance, often outperforming competitors by over 5-10% in accuracy.
  • Plug-and-Play: The method proved effective regardless of which random public dataset was used, confirming that domain-matching is unnecessary.

Limitations and Future Work

  • Meaningless Probes: If the public data is pure noise (e.g., Gaussian noise), the entropy signal becomes meaningless, and SDEA fails to distinguish attackers.
  • Targeted Backdoors: SDEA is designed to protect overall model convergence. It may be less effective against "stealthy" targeted backdoors where the malicious model remains highly confident (sharp) on most data but only fails on specific, rare triggers.
  • Benign Weighting: While the Cooperative Cluster (CC) averages weights for the benign group, "equal weight" might not be the absolute mathematical optimum for every heterogeneous scenario, though it remains the most robust choice.

Summary updated

Abstract

Federated learning presents massive potential for privacy-friendly collaboration. However, the performance of federated learning is deeply affected by byzantine attacks, where malicious clients deliberately upload crafted vicious updates. While various robust aggregations have been proposed to defend against such attacks, they are subject to certain assumptions: homogeneous private data and related proxy datasets. To address these limitations, we propose Self-Driven Entropy Aggregation (SDEA), which leverages the random public dataset to conduct Byzantine-robust aggregation in heterogeneous federated learning. For Byzantine attackers, we observe that benign ones typically present more confident (sharper) predictions than evils on the public dataset. Thus, we highlight benign clients by introducing learnable aggregation weight to minimize the instance-prediction entropy of the global model on the random public dataset. Besides, with inherent data heterogeneity in federated learning, we reveal that it brings heterogeneous sharpness. Specifically, clients are optimized under distinct distribution and thus present fruitful predictive preferences. The learnable aggregation weight blindly allocates high attention to limited ones for sharper predictions, resulting in a biased global model. To alleviate this problem, we encourage the global model to offer diverse predictions via batch-prediction entropy maximization and conduct clustering to equally divide honest weights to accommodate different tendencies. This endows SDEA to detect Byzantine attackers in heterogeneous federated learning. Empirical results demonstrate the effectiveness.
BibTeX
@inproceedings{huang2024self,
  author = {Wenke Huang and Zekun Shi and Mang Ye and He Li and Bo Du},
  title = {{Self-Driven Entropy Aggregation for Byzantine-Robust Heterogeneous Federated Learning}},
  booktitle = {International Conference on Machine Learning},
  year = {2024},
  url = {https://openreview.net/forum?id=k2axqNsVVO},
}
Figure