Decentralized Conformal Novelty Detection via Quantized Model Exchange
Authors: Kyle Loh, Yu Xiang
Summary
The paper addresses novelty detection across a network of independent agents who can't share raw data (privacy or bandwidth constraints). Instead, agents exchange low-precision (quantized) versions of learned scoring functions — think compressed model summaries — and use them to collectively decide what's novel while controlling the global false-discovery rate. The authors prove that evaluating new data against these quantized composite scores maintains the statistical exchangeability property needed for rigorous coverage guarantees, and show empirically that you get competitive detection power with drastically lower communication overhead.
Main takeaways:
- Decentralized novelty detection where agents share quantized scoring functions rather than raw data or full-precision models.
- Maintains finite-sample false-discovery-rate control even with low-precision model exchange, backed by formal proof.
- Drastically cuts communication cost while preserving statistical power in synthetic experiments.
- Could apply to federated or edge settings where bandwidth is limited and privacy matters.
Relevance
Not directly related to my persona/midtraining work — included because it's a useful general-purpose federated-learning technique, potentially relevant if I ever explore distributed fine-tuning or privacy-preserving behavior installation.
Abstract
arXiv:2605.08263v1 Announce Type: new Abstract: This work studies decentralized novelty detection with global false discovery rate (FDR) control across heterogeneous composite null distributions, without sharing the raw data due to privacy and bandwidth considerations. We propose a framework based on the exchange of quantized surrogate models, allowing independent agents to share low-precision representations of locally learned non-conformity score functions. We prove that evaluating data against these quantized composite scores preserves conditional exchangeability, providing rigorous finite-sample guarantees for global FDR control. Empirical studies on synthetic datasets confirm our theoretical results, demonstrating that the proposed approach maintains competitive statistical power while drastically reducing the communication cost.