Research map / Federated learning and privacy / Personalized and heterogeneous FL
Vertical Federated Learning with Missing Features During Training and Inference
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 LASER-VFL, a vertical federated learning method that shares representation and fusion model parameters across predictors for all observed block subsets and trains them via task sampling, handling missing features in both training and inference with convergence guarantees.
Problem
Vertical federated learning standardly requires all feature blocks for every sample during training and inference, so incomplete samples are discarded or predictors become unusable when clients drop out.
Approach
Share representation models across a family of predictors (one per subset of observed blocks containing each client) with average aggregation and a single fusion model per client, and train with unbiased task sampling of subsets so the cost stays linear in K rather than exponential.
Key results
- On CIFAR-100 with K=4 and training/inference missing probability 0.5, LASER-VFL reaches 61.4% accuracy versus 15.8% for Standard VFL and 31.6% for Combinatorial, a 19.3% improvement claimed in the abstract over the baselines.
- With all blocks observed (p_miss=0), LASER-VFL reaches 72.3% on CIFAR-100 versus 64.7% for Standard VFL, and 94.2% on HAPT versus 91.6%.
- Proves O(1/sqrt(T)) convergence for nonconvex L-smooth objectives and linear convergence to a neighborhood of the optimum under the PL inequality.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LASER-VFL rather than Combinatorial | HAPT · HAPT (training p_miss=0.0, inference p_miss=0.0) | accuracy (%): 92.5\pm0.2 -> 94.2\pm0.1 | Helped |
| Use LASER-VFL rather than Combinatorial | HAPT · HAPT (training p_miss=0.5, inference p_miss=0.5) | accuracy (%): 83.0\pm1.3 -> 84.8\pm2.6 | Helped |
| Use LASER-VFL rather than PlugVFL | Credit · Credit (training p_miss=0.0, inference p_miss=0.0) | F1-score × 100: 44.4\pm3.7 -> 46.5\pm2.8 | Helped |
| Use LASER-VFL rather than Ensemble | Credit · Credit (training p_miss=0.5, inference p_miss=0.5) | F1-score × 100: 40.1\pm1.1 -> 41.4\pm1.7 | Helped |
| Use LASER-VFL rather than Standard VFL | MIMIC-IV · MIMIC-IV (training p_miss=0.0, inference p_miss=0.0) | F1-score × 100: 91.6\pm0.2 -> 92.0\pm0.2 | Helped |
| Use LASER-VFL rather than Combinatorial | MIMIC-IV · MIMIC-IV (training p_miss=0.5, inference p_miss=0.5) | F1-score × 100: 78.9\pm0.5 -> 82.0\pm1.0 | Helped |
| Use LASER-VFL rather than Standard VFL | CIFAR-10 · CIFAR-10 (training p_miss=0.0, inference p_miss=0.0) | accuracy (%): 89.7\pm0.2 -> 91.5\pm0.1 | Helped |
| Use LASER-VFL rather than Ensemble | CIFAR-10 · CIFAR-10 (training p_miss=0.5, inference p_miss=0.5) | accuracy (%): 73.8\pm0.6 -> 79.4\pm1.6 | Helped |
| Use LASER-VFL rather than PlugVFL | CIFAR-100 · CIFAR-100 (training p_miss=0.0, inference p_miss=0.0) | accuracy (%): 66.0\pm0.3 -> 72.3\pm0.1 | Helped |
| Use LASER-VFL rather than Ensemble | CIFAR-100 · CIFAR-100 (training p_miss=0.5, inference p_miss=0.5) | accuracy (%): 43.5\pm0.8 -> 51.9\pm1.2 | Helped |
| Use the multi-stage VFedTrans pipeline instead of Local | Credit · Credit (training p_miss=0.0, inference p_miss=0.0) | runtime (s): 127.8 \pm 3.0 -> 2843.0 \pm 50.6 | Hurt |
Surprising findings
- LASER-VFL outperforms standard VFL even when all features are observed, which the authors attribute to a dropout-like regularization effect of task sampling.
- Combinatorial performs better at high test missingness when training missingness is 0.5, because missing test data dictates use of predictors trained on smaller, more frequently observed subsets.
Limitations
- Variance from task sampling grows with K, which enters the convergence neighborhood term.
- Experiments are limited to K=4 clients with feature blocks of images, tabular data and time series; VFedTrans was run only on simpler tasks due to runtime.
- Assumes missing completely at random and that entity alignment identifies observed blocks ahead of training.
Experimental setup
Datasets: HAPT Credit MIMIC-IV CIFAR-10 CIFAR-100
Benchmarks: HAPT Credit MIMIC-IV CIFAR-10 CIFAR-100
Metrics: accuracy F1-score
Baselines: Local Standard VFL VFedTrans Ensemble Combinatorial PlugVFL
Builds on: split neural networks VFedTrans PlugVFL
Closest papers by meaning
- Improved Coresets for Vertical Federated Learning: Regularized Linear and Logistic Regressions (ICML 2025 · similarity 0.87)
- VFLAIR: A Research Library and Benchmark for Vertical Federated Learning (ICLR 2024 · similarity 0.86)
- FedQS: Optimizing Gradient and Model Aggregation for Semi-Asynchronous Federated Learning (NeurIPS 2025 · similarity 0.86)
- FSL-SAGE: Accelerating Federated Split Learning via Smashed Activation Gradient Estimation (ICML 2025 · similarity 0.85)
- Convergence Analysis of Split Federated Learning on Heterogeneous Data (NeurIPS 2024 · similarity 0.85)
- Event-Driven Online Vertical Federated Learning (ICLR 2025 · similarity 0.85)
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.