Kuntech and ETRI Develop On-Device AI Model Obfuscation and Data Flow Analysis Technology

This page was machine-translated and may differ from the original. View original

Kuntech and ETRI Develop On-Device AI Model Obfuscation and Data Flow Analysis Technology

Published2026.09.07 12:43

MLIR Pass-Based Obfuscation and Taint Analysis Technology Integration, Two Related Patents Filed

Kuntech has initiated technology development to protect critical information of model structures and inference code from external analysis in on-device AI environments through a national research project led by ETRI. This research addresses security threats where AI models and applications deployed directly on devices can be subject to internal structure analysis if exposed to attackers.

Kuntech announced on the 7th that in the second year of the 'Development of AI Implementation Information Concealment Technology to Prevent Information Leakage in On-Device AI' project, which is being conducted with support from the Ministry of Science and ICT and the Institute for Information & Communications Technology Planning & Evaluation (IITP), it is developing AI model obfuscation technology and data flow analysis technology based on taint analysis.

The project is led by the Electronics and Telecommunications Research Institute (ETRI) and is being jointly executed by multiple organizations including Kuntech.

The AI model obfuscation technology that Kuntech is developing converts on-device AI models into MLIR (Multi-Level Intermediate Representation) form and then applies obfuscation functions on a pass-by-pass basis.

Key functions include △renaming that changes operator identification information △parameter encapsulation that separates and conceals parameters such as weights and biases △neural structure obfuscation that deforms tensor shapes △graph structure obfuscation that inserts fake connections and nodes, and others.

Kuntech explained that even in obfuscated models, original operations are performed normally by utilizing separate execution information.

The second technology converts C/C++-based AI applications into LLVM IR and tracks the paths through which data is transmitted via variables, pointers, and function calls, using TFLite API as a reference point.

Rather than simple function name patterns, it analyzes actual data dependencies to automatically identify code involved in AI inference.

Kuntech plans to further enhance these two technologies by linking them to selectively protect only critical code related to AI inference in resource-constrained on-device environments.

Two patent applications have been filed for technologies secured during the research process.

Hyuck-joon Bang, CEO of Kuntech, stated, "As on-device AI proliferates, the need to protect not only models but also the core logic of applications that run them is growing," and added, "We will continue to secure on-device AI information protection technology that can be utilized in actual product environments."

#ETRI #임베디드 #쿤텍 #온디바이스 AI #AI 보안 #모델 난독화 #MLIR #오염 분석

To request a correction, reply or follow-up report on this article, see how to file a request. Previously published statements are collected in corrections & replies.