Industrial Internet of Things (IIoTs) serve for efficiency, quality or safety auditing must guarantee data authenticity and integrity by employing access control (AC). While in multi‑collaborative auditing (MCA) IIoTs, the high-privilege access of auditing nodes makes them potential entry points for data leakage, thereby challenging both data integrity and security. This study builds formal model for the MCA‑IIoT setting and derives the latent semantic mapping of abnormal access behaviors involving integrity and security. The derivation shows that semantics‑enhanced data‑driven method together with controllable identity‑masking mechanism is effective against illegal access. On this basis, we propose semantics‑driven and identity‑masking attribute‑based access control (SDIM-ABAC) strategy that merges semantics‑driven federated Bayesian inference to update access thresholds dynamically, while a blockchain-assisted publish–subscribe (Pub/Sub) mechanism supports indirect trusted delivery of raw data. Experiments show that across 22 access-behavior classes, SDIM-ABAC achieves an overall accuracy of 90.3%, with seven anomaly classes detected at 100% accuracy. Compared with state-of-the-art blockchain-based ABACs, SDIM-ABAC reduces time-growth sensitivity under increasing node populations. SDIM‑ABAC improves data security, access compliance and tamper resistance, demonstrating the practicality and scalability in complex collaborative auditing scenarios, it offers a solution for industrial scenarios that must balance sovereign data ownership with collective audit accountability.

Embedding professional knowledge is crucial during developing vertical domain large language models (LLMs). While rapid progress of LLM coexists with barriers to professional knowledge access. Such coexistence highlights the need for effective knowledge management during vertical domain LLM development, yet existing research seldom focuses on this. To solve this gap, we conduct three works: (1) Based on insights of knowledge structural representation, we propose a graph-driven knowledge stream method that models knowledge enabling activities as measurable flows; (2) a knowledge graph (KG)-anchored evaluation protocol is proposed to quantify graph coverage, constraint consistency, hallucination rate, and retention for evaluating knowledge enablement; (3) we integrate graph-anchored evaluation, LLM-graph interaction pipeline, federated learning (FL), and blockchain as graph-driven vertical domain LLM development framework (GVLDF). GVLDF enables knowledge aggregation and reuse, efficacy auditing, and continual optimization. In a foundation pit engineering case study (58,451 tokens fine-tuned vertical domain LLM), GVLDF improved graph coverage from 0.712 to 0.889, raised constraint consistency from 94.3% to 98.1%, and reduced hallucination rate from 0.220 to 0.063, while maintaining retention above 0.87. These findings show that structural anchoring systematically enhances reliability, compliance, and sustainability in vertical LLM development. Our research provides a reliable theory and solution for sustainable enablement and robust deployment of vertical domain LLM.