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

PubSub-VFL: Towards Efficient Two-Party Split Learning in Heterogeneous Environments via Publisher/Subscriber Architecture

NeurIPS 2025 · Federated learning and privacy › Personalized and heterogeneous FL · approach: Vertical federated learning

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 PubSub-VFL, a two-party split learning system combining a Publisher/Subscriber architecture with a parameter server and semi-asynchronous updates, plus a DP-based optimizer for worker and batch-size configuration, reporting 2-7x speedups over VFL baselines.

Problem

Two-party vertical federated learning (split learning) suffers from low computational resource utilization and training latency due to synchronous dependencies between parties, and from resource and data heterogeneity that causes load imbalance.

Approach

Decouple ID alignment from training via batch-ID-indexed embedding and gradient channels in a Pub/Sub architecture, combine with a PS-based intra-party semi-asynchronous mechanism with a tanh-scheduled sync interval, and choose worker counts and batch size by solving a min-max delay optimization solved with dynamic programming over system profiles.

Key results

• PubSub-VFL reports roughly 2-7x training speedup over baselines without accuracy loss, with a 7x running-time reduction versus AVFL-PS on the synthetic dataset.
• Achieves CPU utilization of up to 91.07% and 87.42% at a 50:14 CPU core ratio, versus 42.12% for AVFL-PS in that heterogeneous setting.
• Ablation shows removing the waiting deadline or semi-asynchronous mechanism causes up to 2.10% AUC drop on Synthetic and 1.71% on Bank.

Measured effects reported in the paper

Surprising findings

• Increasing the number of workers or batch size beyond a point slows convergence and reduces efficiency, so more parallelism does not always help.
• Adding DP noise noticeably increases communication cost due to slower convergence, though accuracy and CPU utilization are minimally affected.

Limitations

• Supports only two-party learning; multi-party extension not yet addressed.
• Pipelining effects are approximated with an additive delay model, which may overstate iteration time.
• Experiments use a 64-core CPU server with simulated feature and resource heterogeneity rather than real cross-organization deployment.

Experimental setup

Datasets: Energy Blog Bank Credit Synthetic (scikit-learn, 1M samples, 500 features)

Benchmarks: Energy Blog Bank Credit Synthetic

Metrics: AUC RMSE running time CPU utilization waiting time communication cost attack success rate

Baselines: VFL VFL-PS AVFL AVFL-PS

Builds on: parameter server asynchronous VFL Gaussian differential privacy

Closest papers by meaning

• Federated Transformer: Multi-Party Vertical Federated Learning on Practical Fuzzily Linked Data (NeurIPS 2024 · similarity 0.89)
• VFLAIR: A Research Library and Benchmark for Vertical Federated Learning (ICLR 2024 · similarity 0.88)
• A-FedPD: Aligning Dual-Drift is All Federated Primal-Dual Learning Needs (NeurIPS 2024 · similarity 0.87)
• Integer Is Enough: When Vertical Federated Learning Meets Rounding (AAAI 2024 · similarity 0.87)
• Convergence Analysis of Split Federated Learning on Heterogeneous Data (NeurIPS 2024 · similarity 0.86)
• FedCompetitors: Harmonious Collaboration in Federated Learning with Competing Participants (AAAI 2024 · similarity 0.86)