source: arxiv statistics ml: decentralized conformal novelty detection via quantized model exchange
level: research
this work presents a way to do novelty detection across separate data sites without moving the actual data. each site trains a local model that scores how unusual a data point is. instead of sending raw data or full models, sites share heavily compressed versions of these scoring models. the compression uses quantization, which reduces the model to a low-precision form, saving bandwidth and protecting privacy.
the method guarantees control over the global false discovery rate, meaning the proportion of false alarms among all flagged novelties stays below a set level. this holds even when each site has a different distribution of normal data. the guarantee is finite-sample, so it works for any amount of data, not just in the limit. the key insight is that evaluating data against these quantized scores preserves a property called conditional exchangeability, which underpins the statistical control.
tests on synthetic data show the approach keeps strong detection power while cutting communication costs sharply. the framework suits settings like sensor networks, medical data from multiple hospitals, or any distributed system where raw data cannot be pooled. it offers a practical trade-off: a small loss in precision from quantization in exchange for major savings in data transfer and privacy risk.
why it matters: it enables reliable anomaly detection across distributed data sources without centralizing sensitive information, useful for privacy-sensitive ai applications.
source: arxiv statistics ml: decentralized conformal novelty detection via quantized model exchange