| Technical Name |
LLM-Driven Lightweight AI Intrusion Prevention System for Industrial IoT |
| Project Operator |
National Taiwan University of Science and Technology |
| Project Host |
鄭欣明 |
| Summary |
This technology pioneers an asymmetric AI architecture for IIoT. It uses an LLM to extract the semantics of complex industrial packets, then deploys a lightweight LSTM onto ORing's edge switches for real-time inference. Achieving a 0.99 F1-score, it blocks zero-day attacks without false positives. Additionally, its red teaming engine serves as an automated compliance tool, helping vendors easily pass CRA and SEMI E187 certifications. |
| Scientific Breakthrough |
Our innovation pioneers LLM-based "semantic extraction" for industrial packets, eliminating traditional blind spots. By pairing this with an LSTM Autoencoder, we created an asymmetric, few-shot edge AI architecture achieving a 0.99 F1-score, vastly outperforming current SOTA methods. We also developed an automated red teaming tool to establish dynamic validation standards, forging a comprehensive defense framework. |
| Industrial Applicability |
Resolving the critical "false positive downtime" in OT environments, our technology has integrated with leading networking manufacturer ORing, embedding edge AI directly into industrial switches. Additionally, our red teaming engine functions as an automated compliance testing tool. This helps Taiwanese smart manufacturing and OT equipment vendors rapidly pass CRA and SEMI E187 certifications for global market export. |