Research map / Federated learning and privacy / Personalized and heterogeneous FL

Personalized Subgraph Federated Learning with Differentiable Auxiliary Projections

NeurIPS 2025 · Federated learning and privacy › Personalized and heterogeneous FL · approach: Bayesian personalized FL

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes FedAux, a personalized subgraph federated learning method where each client learns a differentiable auxiliary projection vector whose similarity drives server-side personalized weighted aggregation.

Problem

Personalized subgraph federated learning on non-IID graph data, where the server must estimate client similarity without access to raw data or node embeddings, and high-dimensional parameter comparisons are unreliable.

Approach

Each client learns a 1D auxiliary projection vector (APV) with a differentiable kernel-based soft aggregation over projected node scores (replacing hard sorting + Conv1D), and the server uses cosine similarity between uploaded APVs to compute personalized aggregation weights.

Key results

• FedAux achieves the best accuracy in the main table across all dataset and client-count combinations, e.g., 84.57% on Cora with 5 clients versus 83.72% for FED-PUB.
• On ogbn-arxiv with 5 clients FedAux reaches 68.83% versus 67.62% for FED-PUB.
• Membership inference attack AUC against FedAux ranges 0.49-0.52, near random guessing, versus 0.53-0.58 for baselines.

Measured effects reported in the paper

Limitations

• Theoretical analysis relies on strong assumptions (centred embeddings, linear classifier Jacobian, PL condition), and the Oja equivalence is only approximate for MLP classifiers.
• Results are reported over only three runs in the main table; gains over FED-PUB are often within about 1 point.
• Per-client complexity is quadratic in the number of nodes N_k due to the kernel aggregation.

Experimental setup

Datasets: Cora CiteSeer Pubmed ogbn-arxiv Amazon-Computer Amazon-Photo

Benchmarks: Cora CiteSeer Pubmed ogbn-arxiv Amazon-Computer Amazon-Photo

Metrics: node classification accuracy attack AUC

Baselines: Local FedAvg FedProx FedPer GCFL FedGNN FedGTA FedSage+ FED-PUB

Builds on: FedAvg FED-PUB GCFL Oja learning rule

Closest papers by meaning

• Subgraph Federated Learning for Local Generalization (ICLR 2025 · similarity 0.90)
• Modeling Inter-Intra Heterogeneity for Graph Federated Learning (AAAI 2025 · similarity 0.90)
• Decoupled Subgraph Federated Learning (ICLR 2025 · similarity 0.89)
• FedNE: Surrogate-Assisted Federated Neighbor Embedding for Dimensionality Reduction (NeurIPS 2024 · similarity 0.89)
• FedSSP: Federated Graph Learning with Spectral Knowledge and Personalized Preference (NeurIPS 2024 · similarity 0.89)
• Towards Effective Federated Graph Foundation Model via Mitigating Knowledge Entanglement (NeurIPS 2025 · similarity 0.89)