Security Issues in AI-Centric SoCs - Modeling and Mitigation

Abstract

The computational demands of AI applications require edge devices to be fitted with complex heterogeneous SoCs that may be composed of a variety of hardware accelerators. Such growth in complexity of our systems makes it difficult for hardware and systems designers to ensure these systems are secure against threats, especially novel threats that may arise as a result of heterogeneous design decisions. As AI-centric systems become increasingly pervasive in various application domains, ensuring their security is paramount. This talk explores the range of possible security concerns that can arise in AI-centric SoCs, covering threats at both hardware and software levels. Understanding these vulnerabilities is crucial for developing effective defense strategies. The second part of this talk then takes a deeper dive through a hardware-level anomalous activity detection system that is broadly applicable to heterogeneous SoCs and can enable development of AI-centric SoCs with built-in security. Finally, the talk will motivate future work to develop and deploy security solutions to improve the safety, security, and privacy of modern AI-centric systems.