Geometry-aware federated learning

Current

My current work addresses geometry-aware federated learning for motor-imagery EEG classification. I study federated training of SPDNet on covariance-based representations of EEG signals and compare it with Euclidean baselines such as EEGNet, with a particular emphasis on Riemannian aggregation schemes that preserve the Stiefel/SPD manifold geometry. A central question is how to combine aggregation, partial client participation, and privacy guarantees while remaining compatible with the underlying manifold structure.

θt θt(1) θt(2) θt(3) Server θt → θt+1 Cl. 1 D(1) Cl. 2 D(2) Cl. 3 D(3) Cl. 4 D(4) Cl. 5 D(5)
One communication round: the server broadcasts the global model θt to a sampled subset of clients (dashed clients sit this round out), each updates it locally on private data, and the returning iterates are aggregated into θt+1.
Institution L2S, CentraleSupélec (Gif-sur-Yvette), since October 2025
Supervision Florent Bouchard, with Guillaume Ginolhac and Ammar Mian
Keywords federated learning Riemannian manifolds SPD matrices deep learning

Publications

  1. Preprint T. Pautrel, F. Bouchard, G. Ginolhac, A. Mian Projection-Based Federated Learning on Compact Submanifolds with Geometry-Independent Differential Privacy
    Preprint (PDF)
  2. Accepted T. Pautrel, F. Bouchard, G. Ginolhac, A. Mian FedSPDNet: Geometry-Aware Federated Deep Learning with SPDNet EUSIPCO 2026
    Preprint (PDF) Poster (PDF)

Talks

  1. Federated Learning on Riemannian Manifolds: a Projection-based Approach 7th Statistical Learning for Signal and Image Processing (SLSIP) Workshop, Kruje, Albania
    Slides (PDF)

Earlier work on this theme

Internship · 2023

Riemannian federated learning for time-series analysis with remote sensing data

During an internship at LISTIC, I worked on an early version of FedSPDNet: a geometry-aware federated learning framework targeting the distributed and sovereignty-constrained setting of remote-sensing data. Rather than relying on large convolutional architectures, FedSPDNet operates on Symmetric Positive Definite matrices through second-order statistics, and uses specialized Riemannian aggregation schemes (Stiefel projected means, tangent-space averaging) to preserve the geometric structure of model parameters while reducing communication costs.

Institution LISTIC laboratory, University Savoie Mont Blanc (Annecy)
Supervisors Guillaume Ginolhac, Ammar Mian, Florent Bouchard