A Study on White-Box Cryptography based Integrity Verification Techniques for On-Device AI Model and Their Performance

Abstract

This paper studies White-Box Cryptography (WBC)–based integrity verification for protecting on-device AI models deployed on resource-constrained and potentially hostile platforms. Conventional hash- or MAC-based integrity checks assume that keys and verification logic are isolated from adversaries, an assumption that fails when attackers control the execution environment and can inspect or modify AI binaries and models. WBC embeds secret keys inside heavily obfuscated implementations, making integrity checks more resilient against static and dynamic analysis and key extraction. In this study, methods of using WBC, which offers such security advantages, are investigated to protect on-device AI embedded within embedded systems. To assess the feasibility of WBC-based protection techniques for on-device AI, existing WBC-based integrity verification approaches are selected, implemented, and tested in constrained benchmark environments. Using the algorithmic analysis and experimental results, functional and non-functional requirements are then derived.