Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
139 tokens/sec
GPT-4o
7 tokens/sec
Gemini 2.5 Pro Pro
46 tokens/sec
o3 Pro
4 tokens/sec
GPT-4.1 Pro
38 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

PyVertical: A Vertical Federated Learning Framework for Multi-headed SplitNN (2104.00489v3)

Published 1 Apr 2021 in cs.LG, cs.CR, and cs.DC

Abstract: We introduce PyVertical, a framework supporting vertical federated learning using split neural networks. The proposed framework allows a data scientist to train neural networks on data features vertically partitioned across multiple owners while keeping raw data on an owner's device. To link entities shared across different datasets' partitions, we use Private Set Intersection on IDs associated with data points. To demonstrate the validity of the proposed framework, we present the training of a simple dual-headed split neural network for a MNIST classification task, with data samples vertically distributed across two data owners and a data scientist.

Citations (83)

Summary

  • The paper introduces a novel approach merging split neural networks with Private Set Intersection to enable privacy-preserving vertical federated learning.
  • The framework demonstrates effective training on vertically partitioned data, as evidenced by dual-headed SplitNN experiments on the MNIST dataset.
  • Future research is encouraged to enhance system robustness, scalability, and resistance to adversarial threats in real-world applications.

Overview of PyVertical: A Federated Learning Framework

The paper introduces PyVertical, a framework designed to facilitate Vertical Federated Learning (VFL) using split neural networks. This approach is aimed at enabling the training of neural networks over data that is vertically partitioned across multiple entities, ensuring that raw data remains with its respective owners. This approach is particularly pertinent in scenarios where data privacy is a critical concern, such as financial, health, or other personal datasets. By leveraging Private Set Intersection (PSI) protocols, the framework effectively resolves the challenge of linking data entities across different datasets held by distinct owners without compromising privacy.

Key Components and Methodology

The PyVertical framework integrates two main components: Split Neural Networks (SplitNN) and Private Set Intersection (PSI).

  • Split Neural Networks: In this setup, a neural network is partitioned into segments, each of which is held by a different data owner. These segments transform input data into shareable representations, which are then integrated by a data scientist who may also hold the final segment of the network. This approach encompasses the mapping of data into intermediate, abstract representations allowing for secure and collaborative model training without data owners exposing their raw data.
  • Private Set Intersection: PSI is utilized to identify common data points/shared entities across different datasets. Each data point in a dataset is associated with a unique identifier, and the PSI protocol allows identifying overlaps between datasets without revealing additional information. This mechanism ensures data alignment across parties in a privacy-preserving manner, which is particularly crucial when dealing with sensitive or legally protected data.

Experimental Validation

The framework's validity is demonstrated through an experiment involving a dual-headed SplitNN on the MNIST dataset, which is vertically partitioned between two data owners. The dataset is split such that one owner has the left half of the images, and another owner has the right half. A central data scientist holding the labels coordinates the training process. The experiment demonstrates that despite the data being split across different owners, the proposed setup allows successful training of a neural network, achieving appreciable model performance.

Implications and Future Directions

While the implementation of PyVertical successfully addresses privacy concerns inherent in traditional federated learning setups, its effectiveness is contingent on the honest behavior of participating entities. Future work is necessary to bolster the robustness of such systems against potential threats such as data poisoning or model inversion attacks. Improvements could include the integration of advanced privacy-preserving techniques such as differential privacy or employing decentralized identity verification to ensure authenticated and secure participation processes.

Moreover, while the current framework handles a dual-party scenario, scalability to multiple more complex multi-party datasets remains an open area for exploration. Real-world scenarios often encounter differently sized data segments, necessitating research into handling model segment convergence with varying computational and data contributions. Addressing these challenges could significantly broaden the applicability of PyVertical in various industry domains where data privacy and security are paramount.

In conclusion, PyVertical stands as a promising framework in the domain of Vertical Federated Learning by combining Split Neural Networks with Private Set Intersection to partition and train on private datasets without compromising data confidentiality. This serves as a foundational step towards more efficient, privacy-preserving collaborative machine learning systems. Future research is encouraged to refine and extend its capabilities to accommodate more complex, real-world applications.

Github Logo Streamline Icon: https://streamlinehq.com