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

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

TL;DR

The Federated Learning Sliding Attack (FedSA) introduces a novel class of "controllable" model poisoning attacks by borrowing principles from robust nonlinear control theory. Instead of completely breaking a model, FedSA allows an attacker to precisely dictate a specific level of performance degradation (e.g., dropping global accuracy by exactly 10%) while dynamically adjusting attack speeds to completely evade state-of-the-art Byzantine-robust defenses.

What Problem Does This Paper Address?

Most research on Federated Learning (FL) security focuses on catastrophic poisoning attacks—where malicious actors try to completely break the global model (a Denial-of-Service or DoS attack) or inject a blatant backdoor. Because these attacks cause massive, sudden shifts in model updates, modern Byzantine-robust aggregation rules (AGRs) can easily detect and discard them.

However, a more insidious threat is emerging: controllable poisoning attacks. Imagine a malicious competitor who doesn't want to destroy your service, but instead wants to subtly degrade your model's accuracy by just 5% or 10% to gain a market edge without raising alarms.

Prior attempts at controllable attacks relied on simple model replacement strategies. These approaches suffer from major flaws: they lack theoretical convergence guarantees, require a massive proportion of compromised clients, cannot regulate their poisoning speed, and cannot adapt if the attacker’s goals change mid-training.

Figure 2: The comparison of existing attacks and our attack. The attack effect is illustrated via loss contours—blue area indicates low loss and red area indicates high loss.

Methods

To achieve pinpoint accuracy in model degradation, the authors turn to Sliding Mode Control (SMC), a well-established tool in control engineering used to steer complex, nonlinear dynamic systems along a predefined path.

The authors conceptualize the entire FL optimization process as a nonlinear system. They define an error function ($e_t$) representing the distance between the current global model and a subpar "reference model" (pre-trained to the attacker's target accuracy).

FedSA works via a two-step mechanism:

  1. The Reaching Phase: The attacker manipulates the updates of malicious clients using a custom control law ($u_t$). This forces the global model's trajectory toward a mathematically defined "sliding surface" ($s_t = 0$).
  2. The Sliding Phase: Once on this surface, the global model is locked into a trajectory that guarantees exponential convergence toward the attacker's exact target state.
Figure 3: The mechanism of FedSA.

Crucially, FedSA introduces two hyperparameters that solve the limitations of prior attacks:

  • Control Gain ($k$): Allows the attacker to tune the speed of the attack. By slowing it down, the malicious updates blend seamlessly with normal client variance.
  • Adjustable Variable ($C$): Allows the attacker to dynamically shift the target accuracy mid-attack without needing to generate a new reference model.

Datasets

The authors evaluated FedSA across three standard image classification benchmarks, testing both IID (Independent and Identically Distributed) and Non-IID data splits (using a Dirichlet distribution to simulate real-world data skew):

  • CIFAR10 trained on AlexNet
  • MNIST trained on a Fully Connected (FC) Network
  • Tiny ImageNet trained on ResNet50

Key Insights and Ideas

  • Control Theory Meets ML Optimization: The core "aha moment" is realizing that the iterative aggregation of weights in federated learning behaves exactly like a discrete nonlinear dynamic system, making it highly susceptible to control-theoretic manipulation.
  • Stealth via Trajectory, Not Just Magnitude: Traditional defenses look for outliers in update magnitudes. FedSA achieves stealth by controlling the trajectory of the global model over time, making the degradation look like a natural, slightly sub-optimal training run.
  • Bypassing the Defense Paradox: Because FedSA can precisely throttle its attack speed ($k$), it can successfully pass through strict distance-based and statistics-based defenses (like Krum or Trimmed Mean) that easily catch faster, more aggressive attacks.

Key Results

FedSA was tested against nine different Byzantine-robust defense algorithms (including FedAvg, Median, Trimmed Mean, Bulyan, Mkrum, and FLTrust).

  • Unmatched Precision: Across almost all scenarios, the difference between the attacker's target accuracy and the actual poisoned model accuracy ($|\delta|$) was remarkably small—often under 1% to 2%. For instance, on CIFAR10 with a target accuracy of 60%, FedSA hit within 0.40% of the target under standard FedAvg.
  • Superior Stealth: State-of-the-art attacks like FMPA were frequently detected and neutralized by defenses like Bulyan and Mkrum, causing the model to revert to optimal accuracy. FedSA consistently bypassed these defenses with as few as 10% malicious clients.
  • Controllable Convergence Speed: Ablation studies confirmed that increasing the control gain ($k$) directly accelerates how quickly the model reaches the compromised target state, giving attackers explicit control over their timeline.

Limitations and Future Work

The primary drawback of FedSA is its computational complexity. Because the attacker must calculate the derivative of the server's aggregation function (which is often unknown or non-differentiable), they must approximate it using finite differences. This optimization process makes FedSA slower to execute per round compared to simpler attacks; for example, taking roughly 1 hour and 6 minutes to converge on CIFAR10 compared to ~56 minutes for FMPA.

For future work, the authors highlight the urgent need to develop specialized defense mechanisms capable of detecting these subtle, trajectory-controlled attacks. They also plan to explore extending FedSA to asynchronous federated learning environments.

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 Adams and Abbas 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