Research map / Federated learning and privacy / Communication-efficient FL

Event-Driven Online Vertical Federated Learning

ICLR 2025 · Federated learning and privacy › Communication-efficient FL · approach: Local drift correction · 0 citations (OpenAlex)

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 an event-driven online vertical federated learning framework where only event-triggered clients are activated per round, combined with dynamic local regret for non-convex models in non-stationary streams, with an O(T^{3/4}) regret bound and experiments on i-MNIST.

Problem

Online vertical federated learning assumes all clients receive synchronous feature streams, which is unrealistic; real data arrive via events involving only a subset of clients, and non-convex models in non-stationary streams are not handled by existing online VFL.

Approach

Activate only event-relevant clients each round (others passively contribute embeddings), and adapt dynamic local regret (exponentially time-smoothed sliding-window gradient with buffers) to VFL, with a non-convex regret analysis under random client activation probabilities.

Key results

• Under non-stationary streams, DLR-Random achieves accumulated error 0.0561 vs OGD-Random 0.0788 and SLR-Random 0.1159, with communication ~1465 MB vs SLR ~58589 MB.
• DLR-Event reduces client communication to 1413.52 MB and computation to 2698.96 s versus DLR-Full's 1953.12 MB and 4558.20 s, at accumulated error 0.0681 versus 0.0553.
• Theory gives a non-convex dynamic local regret bound of O(T^{3/4}) with l = T^{1/2} and eta = T^{-1/4}.

Measured effects reported in the paper

Surprising findings

• Under stationary streams, OGD-Event's accumulated error rose from 0.0575 (Full) to 0.1027 with partial activation, while DLR remained near 0.06.
• SLR required communicating its entire buffer each round, giving communication roughly an order of magnitude higher than DLR and OGD.

Limitations

• Passive clients still must participate each round to supply embeddings (incomplete view problem), so communication is not fully eliminated.
• Experiments are limited to a single i-MNIST setup with four clients and small MLPs; the non-stationary stream is synthetic.
• Regret analysis relies on bounded gradient and bounded loss assumptions that are strong for non-convex models.

Experimental setup

Datasets: Infinite MNIST (i-MNIST)

Benchmarks: i-MNIST

Metrics: run-time (prequential) error rate accumulated error rate client computation time communication cost (MB)

Baselines: OGD-Full OGD-Random OGD-Event SLR-Full SLR-Random SLR-Event DLR-Full

Builds on: dynamic local regret online gradient descent static local regret

Closest papers by meaning

• Vertical Federated Learning with Missing Features During Training and Inference (ICLR 2025 · similarity 0.85)
• VFLAIR: A Research Library and Benchmark for Vertical Federated Learning (ICLR 2024 · similarity 0.84)
• Uncertainty-Based Extensible Codebook for Discrete Federated Learning in Heterogeneous Data Silos (ICML 2025 · similarity 0.84)
• Improved Coresets for Vertical Federated Learning: Regularized Linear and Logistic Regressions (ICML 2025 · similarity 0.84)
• On the Necessity of Collaboration for Online Model Selection with Decentralized Data (NeurIPS 2024 · similarity 0.83)
• FedRSPO+: A Heterogeneity-aware Algorithm for Decision-focused Federated Learning (NeurIPS 2026 · similarity 0.82)