Abstract
Federated Learning (FL) is designed to preserve data privacy by sharing model updates instead of raw data. However, Gradient Inversion Attacks (GIAs) have demonstrated that private training data can be reconstructed from these updates. Standard FL protocols like FedAvg involve multiple local training steps ($T > 1$), which aggregate gradient information and challenge existing attacks that rely on linear trajectory assumptions. This paper introduces NL-SME (Non-Linear Surrogate Model Extension), a framework that replaces oversimplified linear models with learnable quadratic Bézier curves. By capturing the non-linear path of model parameters during local training, NL-SME achieves a 94%–98% improvement in gradient matching accuracy over state-of-the-art baselines, revealing significant privacy vulnerabilities in multi-step FL scenarios.
Introduction
Federated Learning (FL) typically employs the FedAvg protocol to minimize communication costs. In this setting, clients perform multiple local optimization steps before sending an aggregated update to the server. This process effectively "blurs" the gradient information, making data reconstruction difficult. Previous state-of-the-art attacks, such as SME, attempted to overcome this by assuming model parameters move in a straight line (linear interpolation) between updates.
However, neural network loss landscapes are inherently non-convex, meaning the actual optimization trajectory is curved. Linear models encounter a "convergence barrier," failing to accurately capture the cumulative effects of multi-step training. NL-SME addresses this by modeling the client's optimization path as a non-linear trajectory, thereby extracting more information from the aggregated gradient than previously thought possible.
Methodology
The core of NL-SME is the use of quadratic Bézier curves to approximate the client's training path. Unlike linear models, a quadratic Bézier curve utilizes a learnable control point ($P_1$) to introduce curvature, allowing the surrogate model to adapt to the actual "bend" in the optimization trajectory.
Joint Optimization
The attack employs a multi-variable joint optimization process. It simultaneously optimizes:
- Dummy Data ($\tilde{D}$): The reconstructed images.
- Curve Position ($t$): The specific point along the trajectory.
- Control Point ($P_1$): The parameter defining the curve's shape.
- Scaling Factors ($d$): Parameter-wise adjustments to account for gradient magnitude variations.
To maintain stability in high-dimensional spaces, the framework incorporates total variation (TV) regularization for image smoothness and specific penalties to keep control points and scaling factors within plausible bounds. A hierarchical learning rate strategy is used to manage the different scales of these variables.
Experimental Results
The researchers evaluated NL-SME on standard benchmarks, including FEMNIST (grayscale characters) and CIFAR-100 (complex RGB images), focusing on small-scale client scenarios ($N \leq 50$).
Key Findings
- The Linear Barrier: Mathematical analysis confirms that linear interpolation incurs an irreducible error proportional to the trajectory's curvature. NL-SME bypasses this bias by modeling the curvature directly.
- Performance Leap: NL-SME demonstrated a 94%–98% improvement in gradient matching accuracy compared to linear baselines like IG and FEDLEAK. Visual reconstructions remained sharp and recognizable, whereas linear methods often produced blurry artifacts.
- Robustness to Heterogeneity: The advantage of NL-SME increases in Non-IID (heterogeneous) settings. In these complex environments, optimization paths become more irregular, making non-linear modeling essential for successful inversion.
Conclusion and Future Work
NL-SME exposes a critical vulnerability in multi-step Federated Learning by demonstrating that curvature in the optimization path is a significant source of privacy leakage. While highly effective under the "honest-but-curious" server assumption, future work could explore active attacks or higher-order curves (e.g., cubic splines) for even greater reconstruction fidelity. These insights serve as a foundation for developing more robust defenses against non-linear trajectory modeling in FL.
Summary updated