Genivara‑PII: Ultra‑Lightweight Browser‑Native Multilingual PII Detection via Knowledge Distillation and Extreme Vocabulary Pruning

This record contains the preprint for Genivara‑PII, an ultra‑lightweight, browser‑native model family for multilingual Personally Identifiable Information (PII) detection. Genivara‑PII addresses the “privacy paradox” by enabling fully local inference inside the browser’s WebAssembly (WASM) sandbox via ONNX Runtime Web, eliminating all server calls for PII detection. Through task‑specific knowledge distillation, extreme vocabulary pruning (from 119,547 to 20,000 tokens), and INT8 quantization, we compress a multilingual DistilBERT teacher (511 MB on disk, 134M FP32 parameters) into an 8.2 MB “nano” model. Despite this 98.4% size reduction, the model attains an aggregate pooled AUROC of 0.9928 (unweighted per‑language mean 0.9706) with a median WASM inference latency of 15.28 ms in the browser. We analyze a high‑trust operating configuration that trades recall (0.57) for near‑zero false positive rate in natural language contexts such as Wikipedia and social media, using an Uncertainty Plateau mechanism and a Hybrid Consistency strategy to suppress uncertain predictions. The approach is directly applicable to enterprise and public‑sector deployments requiring GDPR/CCPA/HIPAA‑aligned processing without cloud data processors, and to air‑gapped or government environments where sensitive data must not leave controlled endpoints; the 8.2 MB footprint and sub‑20 ms latency enable offline deployment on standard end‑user hardware.