CVE/results/paper_method.md

Statistical Method

We constructed a live candidate pool by querying the Censys Platform for 11 IoT device classes and iteratively expanding underrepresented classes until the pool contained 5,800 cross-class unique public IP addresses. IoTProber identified a type label and vendor label from the 46-column multi-perspective fingerprint evidence. We retained candidates only when the Shodan host query returned ok and the IoTProber vendor label was not Unknown, then selected the first 200 eligible addresses per class (2,200 instances in total). Consequently, 100% Shodan host coverage and 100% IoTProber vendor-label availability are sampling constraints and must not be interpreted as unbiased population-level performance estimates.

For every validation IP, we queried Censys and Shodan and applied a native-first mapping hierarchy. Platform-reported CVEs were used first, followed by platform-reported CPEs. Only when neither was available did we construct a type-compatible CPE from that method's own device type, vendor, product, and version observations; baseline methods never reused the IoTProber vendor label. Candidate mappings were restricted to the CISA Known Exploited Vulnerabilities catalog, and explicit product versions were validated against affected-version ranges from CVE.org.

We report pooled and per-device-class coverage on the balanced validation set. Binary proportions include Wilson 95% confidence intervals. Since all methods were evaluated on the same IPs, pairwise differences were tested using the two-sided exact McNemar test and are accompanied by absolute percentage-point differences. Platform-native CVE reports, product-level candidates, and version-confirmed mappings are reported separately.

For the shared vendor metric, we report agreement between each method's device vendor and the IoTProber vendor label. Generic software component vendors such as Apache or Linux are not counted as device-vendor identifications.