Rethinking PUF Design for Scalable Edge AI: A Position on Balancing ML-Attack Resistance and Real-World Deployment

Generative and embedded AI are rapidly migrating from centralized cloud infrastructures to resource-constrained edge devices. While this shift promises reduced latency and improved data privacy, it also creates challenging security and scalability trade-offs. Physical Unclonable Functions (PUFs) are widely touted as low-overhead hardware security primitives suitable for edge and IoT scenarios, yet most existing work emphasizes resistance to machine learning (ML) attacks at the expense of authorized modelability—the ability for trusted entities to accurately model PUF behavior without storing massive Challenge-Response Pair (CRP) databases. This position paper argues that “authorized modelability” should become one of the first-class design objectives for future PUFs. We review existing insights and propose guidelines aimed at balancing ML-attack resistance with the practical requirements of large-scale deployment, thereby addressing a critical yet underexplored aspect of hardware authentication for edge AI.