Let’s Talk Technical: Edge AI Part 3, Security, Standards and Real-World Challenges | DigiKey

What are the biggest challenges developers face as AI moves into embedded and edge devices? In this third and final discussion in DigiKey's Let's Talk Technical series on edge AI, industry experts from NXP Semiconductors, STMicroelectronics, Microchip Technology, Silicon Labs, and Nordic Semiconductor, explore the practical issues shaping the development and deployment of edge AI systems.

The conversation looks beyond the excitement surrounding AI to examine the engineering, security, supply chain, and standards challenges developers are encountering as AI becomes part of embedded products. One major topic is AI security in embedded and edge systems. The discussion explores what security actually means for an AI model, including concerns around model tampering, model spoofing, adversarial behavior, operational certification, and secure software updates. The panel considers how traditional embedded security practices intersect with newer AI-specific threats and why protecting an edge AI system may require expertise across multiple disciplines.

The panel also discusses runtime security and adversarial behavior, including whether available telemetry and indicators can reliably identify malicious or unexpected behavior while an AI system is operating. As AI becomes integrated into products and systems with real-world consequences, understanding how to detect and respond to these threats becomes increasingly important.

Hardware and supply considerations are another part of the conversation. The panel discusses the growing importance of memory for AI workloads, including systems that may require substantial amounts of DDR memory. They also examine the limitations of using headline TOPS (trillions of operations per second) figures to evaluate AI hardware. Real-world inference performance, workload-specific benchmarks, and performance per watt can provide a different picture from theoretical specifications alone.

The discussion also looks at the role of secure hardware, including secure enclaves, communication security gateways, and hardware-based approaches to cryptography. As embedded AI products become more connected and capable, these technologies can become part of a broader security architecture rather than an isolated feature.

Software security is another important consideration. The panel discusses over-the-air (OTA) updates and how security features built into embedded software and operating-system layers can become an important part of protecting AI-enabled products after deployment. This highlights the need to consider security throughout the product lifecycle rather than focusing only on the AI model itself.

Finally, the panel considers whether the edge AI industry needs additional standards and regulatory frameworks. Existing open-source frameworks, connectivity standards, automotive safety practices, cybersecurity requirements, and emerging regulations may all influence how AI systems are developed and deployed. The participants discuss the challenges of establishing standards in a rapidly evolving field where technologies, frameworks, and best practices are still developing.