Title: veriFIRE: An Industrial Case Study in Verifying Consistency Properties for a DNN-Based Wildfire Detection System

Abstract

This paper details the progress of the veriFIRE initiative, a joint effort between academic researchers and industry partners dedicated to enhancing the trustworthiness of real-world, safety-critical systems through formal verification. Our focus is on an airborne platform designed for wildfire detection, which utilizes two deep neural networks. We introduce a comprehensive methodology for verifying consistency properties within this system. The core of our approach involves translating practical, application-specific requirements into queries that are compatible with existing neural network verification solvers.

We examine specific properties across crucial operational scenarios: (i) ensuring that the detector’s confidence increases monotonically as the target intensity rises, and (ii) confirming that the detector’s response remains bounded despite physically realistic sensor blur. By leveraging state-of-the-art neural network verification backends, we implement these encodings and conduct large-scale evaluations using real background data. The results indicate that the first property can be verified efficiently, with all queries resolved in less than five minutes. In contrast, verifying the second property proves significantly more difficult, underscoring the scalability hurdles associated with more complex, high-dimensional specifications. Ultimately, these findings illustrate that it is feasible to derive meaningful, domain-specific guarantees for industrial applications.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC