Sentinel-Bench: A 14-Dimensional Unified Benchmark Standard for Hardware-Aware, Closed-Loop CPS Active Defense

While computer vision and natural language processing have matured around standardized evaluation benchmarks (e.g., ImageNet, GLUE), research in machine learning-based Cyber-Physical System (CPS) security and Network Intrusion Detection Systems (NIDS) remains deeply fragmented. Most academic studies evaluate models against isolated single-dimensional metrics—primarily classification accuracy or F1-score on static offline CSV files—while completely ignoring live line-rate packet ingestion, CPU microarchitecture cache thrashing, worst-case execution time (WCET) determinism, closed-loop physical plant dynamics, and SCADA parser desynchronization zero-days.To establish a unified, open scientific standard, we introduce Sentinel-Bench, the first 14-dimensional hardware-aware benchmark suite for real-time CPS active defense. Built upon the C++20 and eBPF/XDP sentinel lab evaluation framework, Sentinel-Bench provides a standardized multi-metric evaluation methodology across four core pillars: (1) Hardware Systems, Microarchitecture & Cryptography; (2) Cyber-Physical Physics, Formal Safety & Deception; (3) Protocol Grammar, Stateful Fuzzing & Parser Security; and (4) Continual AI Learning, Drift & Federated Safety. We evaluate four representative CPS and NIDS threat corpora (CIC-IDS-2017 PortScan, UNSW-NB15 CPS, SCADA TRITON, and Industroyer2 IEC-104) across all 14 dimensions on bare-metal 10 GbE hardware. Sentinel-Bench establishes open baseline performance benchmarks: achieving an average detection-to-mitigation latency of 0.84 μs (840 ns) at 10,250,000 packets/sec line-rate, an Extreme Value Theory p99.999% WCET bound of 1.82 μs, a CPU IPC of 2.42, formal Control Barrier Function (h(x) ≥ 0) safety set forward-invariance, and zero catastrophic forgetting (BWT = −0.0042). Sentinel-Bench providesan open, reproducible, 14-dimensional benchmark foundation for the global cybersecurity research community