Abstract
Standard Vertical Federated Learning (VFL) requires all participating organizations to be online every time a model makes a prediction, creating a significant hurdle for real-world deployment. We introduce Active-Passive Federated Learning (APFed), a framework where "passive" clients assist in training a robust model but are no longer required during the inference phase. This allows the "active" client to make predictions independently and reliably, overcoming the fragility of traditional VFL systems that break if a partner suffers a network outage or leaves the federation.
Introduction
Vertical Federated Learning (VFL) is designed for scenarios where different organizations hold different features (views) of the same set of users—for example, a bank holding credit scores while a retail platform holds shopping habits. The primary limitation of current VFL methods is their inference strategy: training together and inferring together. To generate a single prediction, every client must process their local data and send representations to a central server. If one partner becomes unavailable months after the model is trained, the entire system fails. APFed addresses this by enabling a training together but inferring solely strategy.
Methods
The APFed framework distinguishes between two types of participants: the Active Client, which initiates the task, holds labels, and owns the final model; and Passive Clients, which provide additional data views during training only.
Training Workflow
- Forward Pass: The active client generates a latent representation (embedding) of its data and sends it to the passive clients.
- Passive Regularization: Passive clients use these embeddings for auxiliary tasks to "guide" the active client's learning. The authors propose two methods:
- APFed-R (Reconstruction): Passive clients attempt to reconstruct their local features from the active client's embedding.
- APFed-C (Contrastive): Passive clients use contrastive learning to align the active client's embeddings with their local feature representations.
- Back-propagation: Passive clients send gradients back to the active client, effectively distilling their cross-view knowledge into the active client's encoder.
Inference Workflow
Once training is complete, the active client retains its optimized encoder and task module. It can generate predictions using only its local data, completely bypassing the passive clients and ensuring system robustness.
Experimental Setup
The framework was evaluated using four standard image benchmarks modified into multi-view settings by splitting images horizontally. This simulates scenarios where different clients see different parts of the same data object.
Datasets included MNIST, Fashion MNIST, CIFAR-10, and CIFAR-100. The researchers tested various split configurations, such as a "2-1 split" where the top half of an image is active and the bottom half is passive.
Key Results and Insights
- Performance Gains: In all tests, APFed significantly outperformed single-view learning (training only on the active client's data). For instance, on CIFAR-100, APFed-C achieved 23.30% accuracy compared to 20.45% for single-view learning.
- Inference Robustness: When passive clients were removed at inference time, standard VFL performance crashed to near-random levels (e.g., ~10% on MNIST). In contrast, APFed maintained its high performance because it was architected for solo inference.
- Contrastive Advantage: APFed-C generally outperformed APFed-R, suggesting that learning high-level feature alignment is more effective for complex data than raw pixel reconstruction.
- Blackbox Flexibility: Because passive clients only exchange embeddings and gradients, they can utilize any local architecture, allowing diverse organizations to participate without revealing internal model structures.
Conclusion and Future Work
APFed provides a flexible and robust alternative to traditional VFL by decoupling the inference phase from passive participants. While currently validated on image classification, future work will explore its application in object detection, time-series analysis, and anomaly detection. Additionally, the authors aim to develop mechanisms to quantify the specific contributions of passive clients to facilitate fair compensation in data marketplaces.
Summary updated