Research map / Federated learning and privacy / Personalized and heterogeneous FL
FedNE: Surrogate-Assisted Federated Neighbor Embedding for Dimensionality Reduction
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 FedNE, a FedAvg-based federated neighbor embedding method that uses learned surrogate repulsion models shared between clients plus intra-client data mixing to visualize distributed data without sharing raw data.
Problem
Training a shared neighbor embedding (dimensionality reduction) model across federated clients is hard because attraction and repulsion terms require pairwise access to data on other clients, so inter-client repulsion is missing and local kNN graphs are incomplete.
Approach
Each client trains a small MLP surrogate that approximates its repulsive loss over 2D grid query points and shares it with other clients; clients also augment local data by mixup-style interpolation between points and their kNN neighbors to address invisible and false neighbors.
Key results
- Averaged over settings, FedNE improves over LocalNE by 2.62% (continuity), 6.12% (trustworthiness), 14.31% (kNN accuracy), 12.69% (steadiness) and 7.31% (cohesiveness).
- On MNIST with Dir(0.1) and 20 clients, FedNE reaches kNN accuracy 0.72 versus 0.48 for FedAvg+NE and trustworthiness 0.85 versus 0.78.
- FedNE adds about 35% GPU time per round over FedAvg on MNIST with 20 clients.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add surrogate repulsion and intra-client data mixing | FedAvg+NE · MNIST, Fashion-MNIST, RNA-Seq, and CIFAR-10 test sets; Dir(0.1), M=20 | kNN accuracy: 0.48 -> 0.72; 0.60 -> 0.66; 0.85 -> 0.90; 0.55 -> 0.63 | Helped |
| add intra-client data mixing (ablation) | FedNE · MNIST and RNA-Seq test sets; Dir(0.1), M=20 | kNN accuracy: 0.65 -> 0.72; 0.87 -> 0.90 | Helped |
| add intra-client data mixing (ablation) | FedNE · MNIST and RNA-Seq test sets; Dir(0.1), M=20 | Trustworthiness: 0.85 -> 0.85; 0.86 -> 0.87 | Mixed |
| add surrogate repulsion model (ablation) | FedNE · MNIST and RNA-Seq test sets; Dir(0.1), M=20 | kNN accuracy: 0.54 -> 0.72; 0.88 -> 0.90 | Helped |
| add surrogate repulsion model (ablation) | FedNE · MNIST and RNA-Seq test sets; Dir(0.1), M=20 | Continuity: 0.96 -> 0.97; 0.97 -> 0.97 | No clear effect |
| increase local kNN graph size from k=7 to k=50 (ablation) | FedNE · MNIST test, Dirichlet(0.1), M=20 | kNN accuracy: 0.72 -> 0.54 | Hurt |
| vary intra-client data mixing weight from α=0.1 to α=0.4 (ablation) | FedNE · MNIST test, Dirichlet(0.1), M=20 | kNN accuracy: 0.72; 0.72; 0.72; 0.72 | No clear effect |
| use grid samples rather than only local embeddings to train the surrogate (ablation) | FedNE · MNIST and Fashion-MNIST test, Shards(2) and Shards(3), M=20 | kNN accuracy: 0.69 -> 0.71; 0.84 -> 0.87; 0.65 -> 0.67; 0.64 -> 0.69 | Helped |
Surprising findings
- On Fashion-MNIST with Dir(0.1), FedNE's kNN accuracy (0.66) is below FedAvg+NE on some metrics, and continuity is essentially tied across methods, so gains are concentrated in trustworthiness, kNN and cluster metrics.
- dSNE and F-dSNE were excluded from comparison because they are non-parametric and rely on a shared public reference dataset.
Limitations
- Experiments use only up to 100 clients and datasets of at most 60K points; scalability to larger data and client counts is not shown.
- Privacy is not formally guaranteed; the surrogate models may still leak information, and privacy techniques like Gaussian mechanisms are only suggested.
- Performance is sensitive to the partition; the paper cannot conclude which component matters most across setups.
Experimental setup
Datasets: MNIST Fashion-MNIST mouse retina scRNA-Seq CIFAR-10
Benchmarks: MNIST Fashion-MNIST scRNA-Seq (RNA-Seq) CIFAR-10
Metrics: continuity trustworthiness kNN classification accuracy steadiness cohesiveness
Baselines: LocalNE FedAvg+NE FedProx+NE GlobalNE
Builds on: FedAvg FedProx contrastive neighbor embedding t-SNE UMAP mixup
Closest papers by meaning
- Personalized Subgraph Federated Learning with Differentiable Auxiliary Projections (NeurIPS 2025 · similarity 0.89)
- Subgraph Federated Learning for Local Generalization (ICLR 2025 · similarity 0.88)
- FedGOG: Federated Graph Out-of-Distribution Generalization with Diffusion Data Exploration and Latent Embedding Decorrelation (AAAI 2025 · similarity 0.88)
- FedMSGL: A Self-Expressive Hypergraph Based Federated Multi-View Learning (AAAI 2025 · similarity 0.87)
- FedRTS: Federated Robust Pruning via Combinatorial Thompson Sampling (NeurIPS 2025 · similarity 0.87)
- Rethinking the Representation in Federated Unsupervised Learning with Non-IID Data (CVPR 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.