Noise-Resilient Plasmonic Physical Unclonable Functions via Deep Contrastive Learning

Physical unclonable functions (PUFs) based on the stochastic optical responses of nanomaterials have emerged as promising hardware security primitives. Their enormous encoding space and inherent randomness produce high-entropy challenge-response characteristics that resist model-based attacks. However, the readout reliability of optical PUFs is susceptible to imaging inconsistencies, including mechanical vibrations and illumination variations. Here, we present a plasmonic PUF system based on spatially disordered gold nanoislands formed by polymer-mediated dewetting. A ConvNeXt-based encoder is trained with combined supervised contrastive, circle, and uniformity losses to achieve an identification accuracy of 99.9%. The resulting embeddings maintain robust discrimination under color perturbations and region-of-interest shifts up to 20%, outperforming traditional direct binarization methods. We further demonstrate the on-chip integration of plasmonic PUFs with both one-time and reusable authentication, highlighting a scalable route toward practical hardware security systems.