📊 statistics

FedSPDnet: Geometry-Aware Federated Deep Learning with SPDnet

FedSPDnet introduces two geometry-aware federated learning frameworks, ProjAvg and RLAvg, designed to preserve the structural integrity of symmetric positive definite (SPD) matrices during aggregation, outperforming standard methods in EEG signal processing tasks.

Original authors: Thibault Pautrel, Florent Bouchard, Ammar Mian, Guillaume Ginolhac

Published 2026-04-27

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Thibault Pautrel, Florent Bouchard, Ammar Mian, Guillaume Ginolhac

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are part of a global team of chefs trying to perfect a secret recipe for a complex sauce. However, there’s a catch: privacy rules prevent you from sharing your actual ingredients or your kitchen setup with anyone else. You can only send "notes" about how your sauce is tasting to a central head chef, who then tries to create a master recipe for everyone.

This paper, FedSPDnet, is about solving a very specific, high-tech version of this "secret recipe" problem.

1. The Problem: The "Square Peg in a Round Hole"

In standard AI (like the "FedAvg" method mentioned), the computer treats all its knowledge like numbers on a flat piece of paper. You can add them, subtract them, and average them easily.

But some types of data—like brain waves (EEG) or radar signals—don't live on a flat piece of paper. They live on curved surfaces, like the surface of a sphere or a donut. In math, these are called Manifolds.

The specific "shape" the researchers are working with is called the Stiefel Manifold. Imagine trying to average the positions of several points on a globe. If you just take the "average" of their coordinates using standard math, your result might end up underground, inside the Earth, where no one can actually walk. In AI, if you "average" these curved parameters using standard methods, the model "breaks"—it loses its geometric structure and stops making sense.

2. The Solution: FedSPDnet (The "Smart Navigator")

The researchers created FedSPDnet. Instead of using "flat" math that breaks the rules of the curved surface, they developed two new ways to "average" the information without falling "underground."

Strategy 1: ProjAvg (The "Snap-Back" Method): Imagine you take the average of several points on a globe, and the result ends up inside the Earth. ProjAvg simply takes that "underground" point and "snaps" it back up to the surface. It’s fast, simple, and keeps the model on track.
Strategy 2: RLAvg (The "Walking the Surface" Method): Instead of going underground and snapping back, this method calculates the direction you would need to walk along the surface to get to the middle. It’s like navigating a mountain range by staying on the trails rather than tunneling through the rock.

3. Why does this matter? (The "Brain Wave" Test)

To prove this works, they tested it on EEG data (brain waves used to control computers).

Brain waves are incredibly complex and are best described using "SPD matrices"—which are essentially mathematical ways of describing the "shape" of brain activity. Because the researchers used a model (SPDnet) that was already designed to understand these shapes, and then used their new "Smart Navigator" (FedSPDnet) to combine the knowledge from different hospitals/users, the results were impressive:

It’s Smarter: It learned much more accurately than standard AI models.
It’s Leaner: It used fewer "parameters" (it’s a smaller, more efficient "recipe").
It’s Tougher: Even when some users didn't participate in every round (partial participation), the model didn't get confused; it kept learning steadily.

Summary in a Nutshell

Most AI tries to solve problems on a flat map. FedSPDnet realizes that some of the most important data in the world (like our brain signals) lives on a curved landscape. By inventing new ways to "average" knowledge while staying on that curve, they’ve made a way for different institutions to collaborate on sensitive data without ever breaking the mathematical rules that make the data meaningful.

Technical Summary: FedSPDnet: Geometry-Aware Federated Deep Learning with SPDnet

1. Problem Statement

Standard Federated Learning (FL) frameworks, such as FedAvg, rely on Euclidean averaging of model parameters. While effective for standard neural networks, this approach fails when model parameters are constrained to non-Euclidean geometries, such as the Stiefel manifold (orthogonal matrices).

The paper specifically addresses the SPDnet architecture, which is designed to process Symmetric Positive Definite (SPD) matrices (e.g., covariance matrices in EEG signals). In SPDnet, the core learnable parameters are orthogonal matrices. Applying standard Euclidean averaging to these weights violates their orthogonality constraints, leading to poor model performance and loss of geometric structure. While Riemannian optimization exists, calculating the true Riemannian mean on the Stiefel manifold is computationally prohibitive for large-scale federated applications.

2. Methodology

The authors propose FedSPDnet, a framework that introduces two lightweight, geometry-preserving aggregation strategies designed to maintain the orthogonality of the Stiefel-constrained weights without the heavy computational cost of true Riemannian means.

A. The SPDnet Architecture
SPDnet uses "BiRe" blocks (BiMap + ReEig) to perform dimensionality reduction while ensuring intermediate representations remain SPD. The weights ( $W$ ) in these blocks belong to the Stiefel manifold ( $St_{p,k}$ ).

B. Proposed Aggregation Strategies
To aggregate local weights $\{W^{(i)}\}_{i \in S_t}$ from $M$ clients, the authors propose:

ProjAvg (Projection Averaging): This method computes the standard arithmetic (Euclidean) mean of the local weights and then projects the result back onto the Stiefel manifold using polar decomposition. This is a computationally efficient way to approximate a manifold-aware mean.
RLAvg (Retraction-Lifting Averaging): This method approximates tangent-space Riemannian averaging. It uses a lifting operation to map the difference between local and global weights into the tangent space, averages them there, and then uses a retraction to map the result back onto the manifold.

Key Advantage: Unlike previous manifold-based FL methods that require clients to use specific Riemannian optimizers, FedSPDnet is optimizer-agnostic. Clients can use any standard optimizer (like Adam or SGD) because the geometric correction happens at the server level during aggregation.

3. Key Contributions

New Federated Framework: The introduction of FedSPDnet, specifically tailored for SPD-preserving deep learning architectures.
Efficient Aggregation Algorithms: The development of ProjAvg and RLAvg, which provide a scalable alternative to the computationally expensive Riemannian mean.
Geometric Preservation: Ensuring that the orthogonality constraints of the Stiefel manifold are maintained throughout the federated training process, preventing "parameter drift" into invalid Euclidean spaces.

4. Results

The authors evaluated FedSPDnet against a Euclidean baseline (FedEEGnet) using two EEG motor imagery datasets: Weibo2014 and PhysionetMI.

Performance: FedSPDnet consistently outperformed FedEEGnet in terms of F1 score. Notably, it retained a higher percentage of its centralized (non-federated) performance compared to the baseline.
Robustness: The model showed high robustness to partial participation (where only half the clients participate in a round) and data heterogeneity.
Efficiency: FedSPDnet achieved better results while using fewer parameters per communication round than the EEGnet baseline.
Convergence: Figures in the paper demonstrate that FedSPDnet converges faster and with lower variance (more stability) across different training seeds.
Algorithmic Equivalence: The experiments showed that ProjAvg and RLAvg perform almost identically, with ProjAvg being slightly preferred due to its lower computational overhead and simpler implementation.

5. Significance

This work is significant for the field of Signal Processing and BCI (Brain-Computer Interface). Since covariance matrices are fundamental descriptors in EEG analysis, the ability to train robust, privacy-preserving models across multiple clinical sites (federated learning) without losing the underlying geometric properties of the data is a major step forward. It bridges the gap between high-level Riemannian geometry and practical, scalable distributed machine learning.