Quantized Model Exchange Enables Decentralized Novelty Detection with FDR Control

A new paper from Kyle Loh and Yu Xiang tackles a critical challenge in distributed systems: how to detect novel events across independent agents without sharing raw data. Their framework, Decentralized Conformal Novelty Detection via Quantized Model Exchange, allows each agent to learn a local non-conformity score function, then share only a low-precision (quantized) version of that model with peers. This approach respects both privacy and bandwidth constraints while enabling global false discovery rate (FDR) control.

The core theoretical contribution is a proof that evaluating data against these quantized composite scores preserves conditional exchangeability. This property is essential for conformal inference methods that require exchangeable data to produce valid prediction sets. By preserving it, the framework provides rigorous finite-sample guarantees for controlling the overall FDR — a key metric when many agents are testing for outliers simultaneously and you want to limit false positives across the entire system.

Empirical tests on synthetic datasets validate the theory. The method maintains competitive statistical power compared to non-quantized alternatives, while drastically reducing communication cost. The authors note that the trade-off between quantization level (bit depth of surrogate models) and detection accuracy is favorable, allowing practitioners to choose a compression rate that fits their network budget.

This work has immediate implications for applications like edge AI, federated anomaly detection, and IoT security, where devices must coordinate to spot rare events without leaking sensitive data. It bridges conformal prediction, distributed optimization, and privacy-preserving machine learning — offering a principled way to scale novelty detection across decentralized networks.