Leakage-Controlled S-Box Fingerprinting Benchmark and Evaluation Artifacts

This dataset provides the CSV-based reproducibility artifacts associated with the manuscript “Leakage-Controlled Interpretable Fingerprinting of Recovered S-Box Tables for Security Triage.” The record supports the reported experiments on generation-family triage of recovered 256-byte bijective S-box tables. The dataset contains the benchmark manifest, model-performance summaries, validation-protocol outputs, feature-block ablation results, random-label and random-permutation negative-control outputs, held-out chaotic-parameter results, metric confidence intervals, per-class GroupKFold metrics, human-readable feature-importance mappings, and the controlled recovered-artifact demonstration result. The benchmark covers four S-box generation families: baseline, RC4-derived, Hexagonal Chaotic Map (HCM), and Logistic Chaotic Map (LCM), with 1,000 unique bijective S-box tables per family. These files are intended to support reproducibility of the leakage-controlled machine-learning evaluation, feature-block interpretation, and recovered-component triage workflow reported in the manuscript. The dataset should be interpreted as a controlled benchmark and evaluation-artifact package for cryptographic component triage. It is not a malware corpus, key-recovery dataset, or universal cipher-identification benchmark.