Lurking in the shadows: Unveiling Stealthy Backdoor Attacks against Personalized Federated Learning

Xiaoting Lyu, Yufei Han, Wei Wang, Jingkai Liu, Yongsheng Zhu, Guangquan Xu, Jiqiang Liu, Xiangliang Zhang

Abstract

Personalized Federated Learning (PFL) is widely regarded as more robust against backdoor attacks than standard Federated Learning (FL). The prevailing assumption is that the local personalization step—where clients fine-tune a global model on their private, clean data—naturally induces "catastrophic forgetting" of any malicious triggers. This paper challenges that assumption by introducing PFedBA (Personalized Federated Backdoor Attack). By mathematically aligning the backdoor task with the model's primary learning objective, PFedBA ensures that the malicious payload survives the personalization process. Our evaluation across diverse datasets and PFL algorithms demonstrates that PFedBA consistently achieves high attack success rates while remaining invisible to state-of-the-art defenses.

Introduction

Federated Learning (FL) has revolutionized collaborative machine learning by allowing clients to train models without sharing raw data. To handle the challenges of non-IID (heterogeneous) data, Personalized Federated Learning (PFL) has emerged as the standard, allowing the global model to be adapted into specialized local versions.

Until now, the "personalization barrier" was thought to be an accidental but effective defense against backdoors. Researchers believed that even if a global model were poisoned, the subsequent local fine-tuning on clean data would erase the backdoor. This paper proves this sense of security is false. We demonstrate that if an attacker aligns the backdoor task to mimic the main task, the personalization process does not erase the backdoor; instead, it helps it "lurk in the shadows," becoming a permanent part of the personalized model.

Methodology: The PFedBA Framework

The core innovation of PFedBA is treating the backdoor as a joint optimization problem rather than a simple data injection task. The attack utilizes two primary mechanisms to ensure persistence and stealth:

1. Gradient Alignment

Drawing on Neural Tangent Kernel (NTK) theory, PFedBA optimizes the backdoor trigger so that its gradient vector aligns with the gradient of the main task. By matching the direction of learning, the malicious updates become statistically indistinguishable from benign ones. This allows the attack to bypass server-side anomaly detection mechanisms (such as Multi-Krum or Trimmed Mean) that look for "outlier" updates.

2. Loss Alignment

To prevent catastrophic forgetting during the personalization phase, PFedBA forces the backdoor task and the main task into the same "loss basin." This ensures that when a benign client fine-tunes the model on their own data, they inadvertently reinforce the backdoor mapping. Instead of the clean data overwriting the poison, the two tasks are mathematically bound together in the model's weight space.

Experimental Setup

We evaluated PFedBA against 10 state-of-the-art PFL algorithms and 6 defense mechanisms using four diverse datasets:

  • Image Classification: Fashion-MNIST, CIFAR-10, and CIFAR-100.
  • IoT Security: N-BaIoT (network traffic data from nine commercial devices), testing the attack on non-image, tabular data.

To simulate realistic PFL environments, we utilized a Dirichlet distribution to create varying degrees of data heterogeneity among clients, ensuring the results are robust across different levels of data non-IIDness.

Key Results and Insights

The Personalization Paradox

Our results reveal a "Personalization Paradox": the very process intended to improve local model utility can be exploited to cement a backdoor. While baseline attacks saw their success rates drop significantly (often below 47%) during personalization, PFedBA maintained an Attack Success Rate (ASR) exceeding 80%—and frequently reaching 99-100%—across all datasets.

Stealth and Integrity

PFedBA successfully circumvented both server-side defenses (like FLAME and DnC) and client-side mitigation techniques (like Neural Cleanse and NAD). Crucially, the attack maintained high accuracy (ACC) on the primary task for benign users. Because the model's performance on legitimate data does not degrade, the presence of the backdoor remains undetected by both the server and the end-user.

Conclusion and Future Work

PFedBA demonstrates that PFL is not inherently resistant to backdoors. By aligning the malicious objective with the benign one, attackers can bypass the personalization barrier and sophisticated filters. This research highlights a critical vulnerability in current PFL architectures and suggests that future defense research must move toward adaptive defenses and randomized smoothing to mitigate optimized, task-aligned triggers.

While our threat model assumes the adversary controls a small percentage of clients (e.g., 10%), future work should investigate the limits of this alignment strategy with even fewer compromised nodes and explore how active defenders might detect the specific gradient patterns inherent in PFedBA.

Summary updated

BibTeX
@inproceedings{lyu2024lurking,
  author = {Xiaoting Lyu and Yufei Han and Wei Wang and Jingkai Liu and Yongsheng Zhu and Guangquan Xu and Jiqiang Liu and Xiangliang Zhang},
  title = {{Lurking in the shadows: Unveiling Stealthy Backdoor Attacks against Personalized Federated Learning}},
  booktitle = {USENIX Security},
  year = {2024},
  url = {https://openreview.net/forum?id=9yH7EW5nz8},
}
Figure