Women’s safety AI systems may achieve high benchmark accuracy while failing during the zero-interaction window of a real attack - the critical period before a victim can trigger an alert. Existing evaluation protocols rarely test passive-detection robustness under physically proximate adversaries operating without device access. This paper introduces the ZIDR Benchmark, an open and reproducible evaluation framework for passive-detection robustness in safety-critical AI systems. Version 2 expands the original framework paper by adding a formal benchmark evaluation across four reference system profiles and a 14-scenario adversarial test library, establishing baseline Zero-Interaction Detection Rate (ZIDR) values and demonstrating that governance compliance does not substitute for adversarial hardening. Three core research artifacts are released. First, a four-layer threat taxonomy formalized as a benchmark schema maps attack surfaces, attack methods, and adversary access levels across sensing, processing, communication, and response layers. Second, Zero-Interaction Detection Rate (ZIDR) is proposed as a scoring metric for evaluating whether a system successfully triggers an alert under adversary-induced degradation without user interaction, accompanied by a Python CLI specification for reproducible system profiling. Third, a grounded 14-scenario library spanning urban and rural Indian deployment contexts is structured for benchmark evaluation across multiple deployment conditions, including urban transit confinement, rural connectivity failure, and sociotechnical suppression scenarios. A standards coverage analysis across the EU AI Act, NIST AI RMF, ISO 42001, India DPDP Act 2023, and the India IT Act identifies the absence of passive-detection robustness requirements and provides a regulatory pathway for operationalizing ZIDR-based conformity assessment. This deposit includes the paper PDF, benchmark schema (YAML), full scenario library, benchmark evaluation outputs, reference system profile definitions, and probe tool specification. A scoring engine implementation, production-grade system profiling templates, and controlled benchmark runs with live vendor systems are in active development and will be released as a subsequent deposit.
Preethi Raghuveeran (Mon,) studied this question.