BROWSE
Detail
Journal Article
온디바이스 AI 정보 유출 위협 동향
- Issue Date
- 2026-10
- Citation
- 전자통신동향분석, v.41, no.5, pp.61-71
- ISSN
- 1225-6455
- Publisher
- 한국전자통신연구원
- Language
- Korean
- Type
- Journal Article
- Abstract
- On-device artificial intelligence (AI) is expanding across mobile, Internet of Things (IoT), and edge computing environments, driven by its advantages in privacy preservation, low latency, and reduced network costs. However, storing and executing AI models alongside user data directly on local devices introduces new information-leakage risks that conventional, cloud-centric security mechanisms fail to fully address. This article reviews representative information-leakage attacks targeting cloud-based AI systems and comprehensively analyzes threat models, attack vectors, real-world case studies, and recent research trends specific to on-device AI environments. It examines model extraction, training-data inference, adversarial attacks, backdoor insertion, and information leakage originating from the underlying execution environments. These threats significantly undermine the security and trustworthiness of on-device AI systems. Recent studies, such as the Reverse Engineering of On-device Models (REOM) framework and Bin-to-DNN, have demonstrated that deployed models can be reconstructed to a degree that enables near-white-box analysis. Concurrently, high-profile incidents involving Microsoft Recall and the LeftoverLocals vulnerability highlight that sensitive user data and model parameters can leak even within commercial production environments. This review identifies the limitations of current protection mechanisms and underscores the critical need for multilayered defenses—including model encryption, Trusted Execution Environments (TEEs), integrity verification, code obfuscation, and digital watermarking. Implementing these countermeasures is vital to fostering the development of secure and resilient on-device AI ecosystems.
- Keyword
- AI Data Leakage, AI Security, Edge AI Security, Mobile AI Security, Model Encryption, Model Extraction, Obfuscation, On-Device AI, TEE
- KSP Keywords
- Adversarial Attacks, Case studies, Cloud-centric, Commercial production, Current protection, Data Leakage, Edge Computing, Information leakage, Integrity verification, Internet of Things(IoT), Low Latency
This work is distributed under the term of Korea Open Government License (KOGL)
(Type 4: : Type 1 + Commercial Use Prohibition+Change Prohibition)
(Type 4: : Type 1 + Commercial Use Prohibition+Change Prohibition)
