Exploit Gradient Skewness to Circumvent Byzantine Defenses for Federated Learning

Yuchen Liu, Chen Chen, Lingjuan Lyu, Yaochu Jin

TL;DR

This paper introduces a novel Byzantine attack called STRIKE that exploits a newly discovered phenomenon called "gradient skew" in federated learning. By hiding malicious gradients within skewed honest gradients, STRIKE effectively bypasses existing defenses that rely on the assumption that dense gradients are always honest.

What Problem Does This Paper Address?

Federated Learning (FL) allows multiple clients to collaboratively train a machine learning model without sharing their private data. However, FL is vulnerable to Byzantine attacks, where a subset of malicious clients can send manipulated gradients to corrupt the global model. Many existing defenses aim to identify and discard these malicious gradients by assuming that the majority of "dense" gradients are honest. This paper addresses the critical gap that these defenses are susceptible to attacks that exploit the non-IID (non-independent and identically distributed) nature of data, which leads to a phenomenon called "gradient skew." This skew causes honest gradients to deviate from the true optimal gradient, creating a blind spot for defenses that trust dense gradients.

Methods

The core of the paper is the STRIKE attack, a two-stage process designed to exploit gradient skew:

  1. Stage 1: Search for Skewed Honest Gradients: The attack identifies "skewed gradients" by leveraging Karl Pearson's formula. It searches for gradients that lie in the direction from the mean of honest gradients towards the median of honest gradients. Gradients with the highest projection onto this search direction are identified as skewed honest gradients.
  2. Stage 2: Hide Byzantine Gradients: The attack then constructs Byzantine gradients that are placed within the identified skewed honest gradients. This is achieved by solving a constrained optimization problem that aims to maximize the distance between the aggregated malicious gradients and the true optimal gradient, while ensuring the Byzantine gradients remain within the "diameter" of the skewed honest gradients. This makes the malicious gradients appear as part of the dense, seemingly honest, skewed gradient cluster.

Datasets

The authors evaluated their attack on three benchmark datasets:

  • CIFAR-10: A widely used image classification dataset.
  • ImageNet-12: A subset of the ImageNet dataset, used for its challenging nature.
  • FEMNIST: A dataset derived from the MNIST handwritten digit dataset, adapted for federated learning scenarios.

The experiments were conducted using non-IID data distributions generated via the Dirichlet distribution, a common method for simulating heterogeneous data in FL.

Key Insights and Ideas

  • Gradient Skew is a Real Phenomenon: The paper's primary contribution is the discovery and demonstration of "gradient skew" in FL with non-IID data. This phenomenon, where dense honest gradients deviate from the optimal gradient, is crucial.
  • Inductive Bias is a Weakness: Existing Byzantine defenses often rely on the assumption that dense gradients are honest. The paper shows how this assumption becomes a vulnerability when honest gradients themselves are skewed.
  • Exploiting the Blind Spot: STRIKE effectively exploits this blind spot by making Byzantine gradients indistinguishable from the skewed honest gradients, thus fooling defenses.
  • Outlier Gradients are Important: The attack's success also highlights that by hiding within skewed gradients, Byzantine clients can effectively mask themselves, potentially leading to the misclassification or removal of genuinely useful outlier honest gradients, which can harm model generalization.

Key Results

  • Superior Attack Performance: STRIKE consistently outperformed several state-of-the-art Byzantine attacks (BitFlip, LIE, IPM, MinMax, MinSum, Mimic) across various defenses (Multi-Krum, Median, RFA, Aksel, CClip, DnC, RBTM) on CIFAR-10, ImageNet-12, and FEMNIST.
  • Significant Improvement Against DnC: STRIKE showed a notable improvement of 57.84% against the DnC defense on FEMNIST with 20% Byzantine clients.
  • Effectiveness on Heterogeneous Data: The attack's effectiveness was particularly pronounced on datasets with higher skew (ImageNet-12 and FEMNIST), suggesting that the degree of non-IIDness directly impacts the attack's success.
  • Robustness Across Settings: STRIKE demonstrated effectiveness across varying Byzantine client ratios and different levels of non-IID data.

Limitations and Future Work

  • IID Data Limitation: The authors acknowledge that the effectiveness of STRIKE is closely tied to the gradient skew phenomenon, which is prevalent in non-IID settings. In IID settings, the attack's performance might be limited.
  • Defense Mitigation: Defenses that can effectively alleviate non-IID data distributions could potentially mitigate the STRIKE attack.
  • Future Work: The authors suggest exploring potential defenses against this skew-aware attack as a future research direction. They also hint at exploring the impact of their attack on model generalization due to the potential misclassification of outlier honest gradients.

Summary updated

Abstract

Federated Learning (FL) is notorious for its vulnerability to Byzantine attacks. Most current Byzantine defenses share a common inductive bias: among all the gradients, the densely distributed ones are more likely to be honest. However, such a bias is a poison to Byzantine robustness due to a newly discovered phenomenon in this paper - gradient skew. We discover that a group of densely distributed honest gradients skew away from the optimal gradient (the average of honest gradients) due to heterogeneous data. This gradient skew phenomenon allows Byzantine gradients to hide within the densely distributed skewed gradients. As a result, Byzantine defenses are confused into believing that Byzantine gradients are honest. Motivated by this observation, we propose a novel skew-aware attack called STRIKE: first, we search for the skewed gradients; then, we construct Byzantine gradients within the skewed gradients. Experiments on three benchmark datasets validate the effectiveness of our attack
BibTeX
@inproceedings{liu2025exploit,
  author = {Yuchen Liu and Chen Chen and Lingjuan Lyu and Yaochu Jin and Gang Chen},
  title = {{Exploit Gradient Skewness to Circumvent Byzantine Defenses for Federated Learning}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i18.34094},
}
Figure