[Paper Review] Asymmetrical Vertical Federated Learning

Yang Liu, Xiong Zhang+1 more|arXiv (Cornell University)|Apr 16, 2020

Privacy-Preserving Technologies in Data10 references31 citations

TL;DR

The paper introduces asymmetrical vertical federated learning (AVFL) to protect sample ID privacy, develops an asymmetrical private set intersection (APSI) protocol with a Pohlig-Hellman realization, and proposes a Genuine with Dummy (GWD) training approach demonstrated via federated logistic regression.

ABSTRACT

Federated learning is a distributed machine learning method that aims to preserve the privacy of sample features and labels. In a federated learning system, ID-based sample alignment approaches are usually applied with few efforts made on the protection of ID privacy. In real-life applications, however, the confidentiality of sample IDs, which are the strongest row identifiers, is also drawing much attention from many participants. To relax their privacy concerns about ID privacy, this paper formally proposes the notion of asymmetrical vertical federated learning and illustrates the way to protect sample IDs. The standard private set intersection protocol is adapted to achieve the asymmetrical ID alignment phase in an asymmetrical vertical federated learning system. Correspondingly, a Pohlig-Hellman realization of the adapted protocol is provided. This paper also presents a genuine with dummy approach to achieving asymmetrical federated model training. To illustrate its application, a federated logistic regression algorithm is provided as an example. Experiments are also made for validating the feasibility of this approach.

Motivation & Objective

• Define and formalize asymmetrical vertical federated learning (AVFL) and distinguish it from symmetrical vertical FL.
• Adapt private set intersection (PSI) to protect weak-party IDs during ID alignment.
• Provide a Pohlig-Hellman based APSI realization and analyze its security properties.
• Propose the Genuine with Dummy (GWD) approach for asymmetrical model training.
• Demonstrate AVFL via a federated logistic regression example and validate feasibility with experiments.

Proposed method

• Formally classify vertical FL into symmetrical and asymmetrical based on pre-ID-alignment ID spaces.
• Introduce Asymmetrical PSI (APSI) that yields an obfuscated set obf with private intersection kept by the weak party.
• Provide a Pohlig-Hellman based APSI protocol and prove security properties of obf and intersection.
• Propose a Genuine with Dummy (GWD) training framework with a trusted coordinator to preserve asymmetry while keeping results equivalent to SVFL.
• Apply GWD to asymmetrical vertical logistic regression and detail the encrypted gradient computation and updates.

Experimental results

Research questions

• RQ1 How can vertical federated learning protect the privacy of sample IDs in asymmetrical settings where one party is weaker in ID exposure?
• RQ2 Can APSI obfuscation preserve necessary joint data for training while limiting ID leakage to the weak party?
• RQ3 Does the GWD training approach yield comparable performance to standard symmetrical vertical FL?
• RQ4 How does AVFL perform in a practical model (e.g., logistic regression) under secure computation constraints?

Key findings

• APFL (APSI) enables privacy-preserving ID alignment by obfuscating the intersection while allowing the weak party to learn the true intersection and keeping strong-party privacy.
• A Pohlig-Hellman based APSI protocol securely computes a obfuscated intersection set with tunable privacy parameter λ.
• The Genuine with Dummy (GWD) approach ensures the asymmetrical training yields the same results as standard vertical FL by masking dummy samples with mathematical identities.
• Experimentation on MNIST with AVLR shows training loss and AUC trajectories are similar across different λ values, indicating feasibility and comparable performance to symmetrical setups.