ML Security Papers3D Guard-Layer: An Integrated Agentic AI Safety System for Edge Artificial Intelligence
Eren Kurshan 1, Yuan Xie 2, Paul Franzon 3
Published on arXiv
2511.08842
Input Manipulation Attack
OWASP ML Top 10 — ML01
Excessive Agency
OWASP LLM Top 10 — LLM08
Key Finding
The co-located 3D-integrated safety layer improves resilience against network-based attacks on edge AI systems with minimal hardware overhead.
3D Guard-Layer
Novel technique introduced
AI systems have found a wide range of real-world applications in recent years. The adoption of edge artificial intelligence, embedding AI directly into edge devices, is rapidly growing. Despite the implementation of guardrails and safety mechanisms, security vulnerabilities and challenges have become increasingly prevalent in this domain, posing a significant barrier to the practical deployment and safety of AI systems. This paper proposes an agentic AI safety architecture that leverages 3D to integrate a dedicated safety layer. It introduces an adaptive AI safety infrastructure capable of dynamically learning and mitigating attacks against the AI system. The system leverages the inherent advantages of co-location with the edge computing hardware to continuously monitor, detect and proactively mitigate threats to the AI system. The integration of local processing and learning capabilities enhances resilience against emerging network-based attacks while simultaneously improving system reliability, modularity, and performance, all with minimal cost and 3D integration overhead.
Key Contributions
- 3D chip-stacked hardware integration of a co-located safety monitoring layer with edge AI accelerators
- Adaptive, agent-based threat detection and mitigation system that continuously learns emerging attack patterns
- Safety architecture designed for low-overhead deployment on resource-constrained edge devices
🛡️ Threat Analysis
The safety architecture is specifically designed to monitor, detect, and mitigate attacks against deployed edge AI systems at inference time, including 'network-based attacks' — the primary threat vector described aligns with inference-time input manipulation and evasion attacks on edge AI models.