Performance Guaranteed Poisoning Attacks in Federated Learning: A Sliding Mode Approach

Huazi Pan, Yanjun Zhang, Leo Yu Zhang, Scott D. Adams, Abbas Z. Kouzani, Suiyang Khoo

Introduction

Federated Learning (FL) has emerged as a privacy-preserving paradigm for training machine learning models across decentralized clients. However, its distributed nature makes it susceptible to poisoning attacks, where malicious clients submit crafted updates to manipulate the global model. While traditional attacks often aim for total model failure (untargeted) or specific misclassifications (targeted), a more insidious threat is controllable poisoning. This involves subtly degrading a model's performance to a precise, pre-defined level—a "precision strike" that can provide a competitive advantage while remaining undetected by standard defenses.

The Challenge of Precision Poisoning

Most existing poisoning attacks are "best-effort" strategies. They lack mathematical guarantees for convergence to a specific target accuracy and are often slow or easily detected by Byzantine-robust aggregation rules (AGRs) like Krum, Median, or Trimmed Mean. For instance, the Flexible Model Poisoning Attack (FMPA) attempts controllable poisoning but often fails to hit the target accurately or converges too slowly.

FedSA: A Sliding Mode Control Approach

To address these limitations, the authors of "Performance Guaranteed Poisoning Attacks in Federated Learning" introduce FedSA (Federated Learning Sliding Attack). FedSA is the first poisoning attack to leverage Sliding Mode Control (SMC), a robust control technique typically used in high-precision engineering fields like robotics.

By treating the FL optimization process as a nonlinear dynamic system, FedSA defines a mathematical "sliding surface" based on the error between the current global model and a target "subpar" reference model. The attack then employs a control law to calculate the exact updates needed to move the global model toward this surface and keep it there.

FedSA targets a specific attack objective on the loss landscape, contrasting with the divergent paths of traditional attacks like LIE or Min-Max.

Key Methodological Innovations:

  1. Guaranteed Convergence: Unlike heuristic attacks, FedSA is mathematically proven to reach its poisoning objective within a finite number of communication rounds.
  2. Stealth through Approximation: To bypass unknown server-side defenses, FedSA uses a finite difference approximation to estimate the server's aggregation behavior, allowing it to remain effective against various AGRs.
  3. Dynamic Flexibility: The attack objective can be adjusted in real-time. By changing a constant in the control law, an attacker can shift the target accuracy mid-training.

Experimental Evaluation

The effectiveness of FedSA was evaluated on three standard benchmarks: MNIST (Fully Connected network), CIFAR-10 (AlexNet), and Tiny ImageNet (ResNet50). These were tested under both IID and Non-IID data distributions to ensure robustness in realistic federated environments.

Key Results and Discussion

FedSA consistently outperformed state-of-the-art (SOTA) attacks across all tested scenarios. The primary metric of success was the precision with which the attack hit the target accuracy.

Test accuracy over time for FedSA compared to FMPA and the "No Attack" baseline. FedSA (red) demonstrates superior precision and convergence to the Target Acc. (green) across different settings.

Performance Highlights:

  • High Precision: FedSA often achieved the target accuracy with an error of less than 1%, whereas previous attacks like FMPA missed by significant margins (10-30%).
  • Efficiency: The attack remains potent even with a small fraction of malicious clients (as low as 10%).
  • Defense Bypassing: FedSA successfully navigated a wide array of Byzantine-robust defenses, including Krum, Bulyan, and FLTrust. The authors noted that FedSA improved attack effectiveness by orders of magnitude compared to traditional methods in certain high-defense scenarios.

Conclusion and Future Directions

FedSA represents a significant shift in the landscape of FL security, moving poisoning from heuristic guessing to a mathematically grounded control problem. Its ability to provide performance guarantees makes it a formidable threat that current defenses are ill-equipped to handle. Future research should focus on developing specialized defenses capable of detecting the subtle, trajectory-based signatures of control-theory-driven attacks.

This paper was presented at IJCAI-25.

Summary updated

Abstract

Manipulation of local training data and local updates, i.e., the poisoning attack, is the main threat arising from the collaborative nature of the federated learning (FL) paradigm. Most existing poisoning attacks aim to manipulate local data/models in a way that causes denial-of-service (DoS) issues. In this paper, we introduce a novel attack method, named Federated Learning Sliding Attack (FedSA) scheme, aiming at precisely introducing the extent of poisoning in a subtle controlled manner. It operates with a predefined objective, such as reducing global model's prediction accuracy by 10%. FedSA integrates robust nonlinear control-Sliding Mode Control (SMC) theory with model poisoning attacks. It can manipulate the updates from malicious clients to drive the global model towards a compromised state, achieving this at a controlled and inconspicuous rate. Additionally, leveraging the robust control properties of FedSA allows precise control over the convergence bounds, enabling the attacker to set the global accuracy of the poisoned model to any desired level. Experimental results demonstrate that FedSA can accurately achieve a predefined global accuracy with fewer malicious clients while maintaining a high level of stealth and adjustable learning rates.
BibTeX
@inproceedings{pan2025performance,
  author = {Huazi Pan and Yanjun Zhang and Leo Yu Zhang and Scott D. Adams and Abbas Z. Kouzani and Suiyang Khoo},
  title = {{Performance Guaranteed Poisoning Attacks in Federated Learning: A Sliding Mode Approach}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.24963/ijcai.2025/670},
}
Figure