Class-Conditional Robust Conformal Prediction for Structured Perturbations

Abstract

We introduce a conformal prediction (CP) method that leverages class-conditional randomized smoothing with spectrally localized perturbations to address structured corruption in AI-driven multispectral image classification. This setting reflects realistic corruptions such as those found in remote sensing and other sensor-based applications, where specific object categories may be disproportionately affected by environmental or hardware-induced fluctuations, e.g. specifically only in red spectral channel or near-infrared channels. The Randomized Smoothed Conformal Prediction (RSCP) framework makes use of global uniform noise to construct valid prediction sets. Since real-world perturbation are rarely uniform, RSCP would lead to an increased prediction set size for uncorrupted classes, reducing informativeness and efficiency of the conformal method. For such asymmetric perturbations, we propose a class-conditional RSCP framework in which perturbations are applied only to certain target classes. Our approach allows for risk-stratified robustness, providing more nuanced uncertainty estimates to critical or noise-prone classes without sacrificing coverage for unaffected categories. Class conditional coverage guarantees for smoothed scores with class-dependent noise is guaranteed.