AXIOS: Privacy-Preserving Vertical Federated Learning for Anomaly Detection
Faculty Mentor Information
Dr. Gaby Dagher, Boise State University; and Dr. Tim Andersen, Boise State University
Presentation Date
7-16-2026
Abstract
As machine learning (ML) becomes increasingly incorporated into daily life, it is crucial that the data and training that define these systems are regulated for quality. Federated Learning (FL) builds on this value by providing a mechanism for collaborative machine learning between cross-domain or cross-silo parties without requiring the sharing of their private data, and Vertical Federated Learning (VFL) particularly enables training across parties with shared samples but partitioned feature spaces. However, the unique challenge of distributing ML training across parties in a VFL setting complicates essential ML model quality-preserving mechanisms such as anomaly detection. In this paper, we introduce AXIOS, a differentially private VFL framework that implements anomaly detection on intermediate training data in order to improve ML model accuracy without sacrificing privacy. We use a majority-based consensus protocol through blockchain in order to decentralize anomaly detection by taking advantage of cross-party feature distributions. We describe the process of utilizing autoencoder reconstruction loss as an anomaly metric and outline the tradeoff between differential privacy budget and anomaly detector accuracy, encouraging future research on the potential for anomaly detection on intermediate ML data.
AXIOS: Privacy-Preserving Vertical Federated Learning for Anomaly Detection
As machine learning (ML) becomes increasingly incorporated into daily life, it is crucial that the data and training that define these systems are regulated for quality. Federated Learning (FL) builds on this value by providing a mechanism for collaborative machine learning between cross-domain or cross-silo parties without requiring the sharing of their private data, and Vertical Federated Learning (VFL) particularly enables training across parties with shared samples but partitioned feature spaces. However, the unique challenge of distributing ML training across parties in a VFL setting complicates essential ML model quality-preserving mechanisms such as anomaly detection. In this paper, we introduce AXIOS, a differentially private VFL framework that implements anomaly detection on intermediate training data in order to improve ML model accuracy without sacrificing privacy. We use a majority-based consensus protocol through blockchain in order to decentralize anomaly detection by taking advantage of cross-party feature distributions. We describe the process of utilizing autoencoder reconstruction loss as an anomaly metric and outline the tradeoff between differential privacy budget and anomaly detector accuracy, encouraging future research on the potential for anomaly detection on intermediate ML data.