重写手机AI安全边界！首个MoAI攻防SoK，系统梳理三大安全支柱
Merged summary
TL;DR - A systematization-of-knowledge paper maps the security landscape of mobile on-device AI, where local inference improves privacy and latency but exposes models and runtimes to direct attack. It provides a unified framework for assessing threats and defenses across the entire device-side AI pipeline.
- Models four system layers: input interfaces, model artifacts, runtime execution, and hardware isolation.
- Defines three security pillars: input integrity, resident-model security, and execution-environment confinement.
- Organizes threats into adversarial examples, backdoors, weight manipulation, model theft, and energy/latency attacks.
- Surveys obfuscation, authorization, trusted execution environments, and watermarking, while identifying open challenges in on-device training, generative AI, and agents.
Sources (1)
重写手机AI安全边界！首个MoAI攻防SoK，系统梳理三大安全支柱
TL;DR - A systematization-of-knowledge paper maps the security landscape of mobile on-device AI, where local inference improves privacy and latency but exposes models and runtimes to direct attack. It provides a unified framework for assessing threats and defenses across the entire device-side AI pipeline.
- Models four system layers: input interfaces, model artifacts, runtime execution, and hardware isolation.
- Defines three security pillars: input integrity, resident-model security, and execution-environment confinement.
- Organizes threats into adversarial examples, backdoors, weight manipulation, model theft, and energy/latency attacks.
- Surveys obfuscation, authorization, trusted execution environments, and watermarking, while identifying open challenges in on-device training, generative AI, and agents.