# Quadratic Functional Encryption for Secure Training in Vertical Federated Learning

**Authors:** Shuangyi Chen, Anuja Modi, Shweta Agrawal, Ashish Khisti  
**arXiv:** [2305.08358](https://arxiv.org/abs/2305.08358)  
**Date:** 2023-05-15  
**Summary page:** https://www.summarizepaper.com/en/paper/2305.08358v1/

## Key points

- Federated learning revolutionized the field of machine learning
- Vertical federated learning (VFL) is a specific type of federated learning where each party holds a subset of features and only one party holds the labels
- Ensuring privacy in VFL remains a challenge
- FedV is a framework for secure gradient computation in VFL using multi-input functional encryption proposed by Xu et al.
- Chen et al. propose using Quadratic functional encryption to avoid information leakage when training generalized linear models for VFL
- Quadratic functional encryption can be used to encrypt intermediate values during the computation process, preventing any individual party from gaining access to sensitive information about other parties' data
- This approach offers significant improvements over previous methods that relied on homomorphic encryption or secure multiparty computation techniques that were computationally expensive and impractical for large-scale datasets
- The authors provide experimental results demonstrating the effectiveness of their approach in terms of both accuracy and privacy preservation

## Summary

The field of machine learning (ML) has been revolutionized by the advent of federated learning, which allows multiple parties to collaborate on training models without sharing their private data. Vertical federated learning (VFL) is a specific type of federated learning where each party holds a subset of the features and only one party holds the labels. However, ensuring privacy in VFL remains a challenge. Recently, Xu et al. proposed FedV, a framework for secure gradient computation in VFL using multi-input functional encryption. In this paper, Chen et al. build upon this work and propose using Quadratic functional encryption to avoid information leakage when training generalized linear models for VFL. The authors explain that Quadratic functional encryption can be used to encrypt intermediate values during the computation process, preventing any individual party from gaining access to sensitive information about other parties' data. This approach offers significant improvements over previous methods that relied on homomorphic encryption or secure multiparty computation techniques that were computationally expensive and impractical for large-scale datasets. The authors provide experimental results demonstrating the effectiveness of their approach in terms of both accuracy and privacy preservation. Overall, this work represents an important step forward in enabling secure collaboration among multiple parties in ML applications while preserving individual data privacy.