Sensitivity of Anomaly Detection to Prior Specification

Determine the sensitivity of the Bayesian inverse-problem framework for unsupervised anomaly detection to the specification of the prior distribution over corruption parameters, through systematic sensitivity analysis.

Background

The framework places a prior distribution on anomaly or corruption parameters, and this prior contributes directly to the anomaly energy used for scoring observations. Although the paper evaluates several corruption functions and priors, it does not systematically study how prior specification affects detection or inference performance. The authors explicitly leave this analysis for future work, making the dependence of the framework’s results on prior choice an unresolved question.