Artificial optical microfingerprints, known as physically unclonable functions (PUFs) offer a groundbreaking approach for anti-counterfeiting. However, these PUFs artificial optical microfingerprints suffer from a limited number of challenge-response pairs, making them vulnerable to machine learning (ML) attacks when additional error-correcting units are introduced. This study presents a pioneering demonstration of artificial optical microfingerprints that combine the advantages of PUFs, a large encoding capacity algorithm, and reliable deep learning authentication against ML attacks. Our approach utilizes the triple-mode PUFs, incorporating bright-field, multicolor fluorescence wrinkles, and the topography of surface enhanced Raman scattering in the mechanical and optical layers. Notably, the quaternary encoding of these PUFs artificial microfingerprints allows for an encoding capacity of 6.43 × 10²⁴⁰⁸² and achieves 100% deep learning recognition accuracy. Furthermore, the PUFs artificial optical microfingerprints exhibit high resilience against ML attacks, facilitated by generative adversarial networks (GAN) (with mean prediction accuracy of ~ 85.0%). The results of this study highlight the potential of utilizing up to three PUFs in conjunction with a GAN training system, paving the way for achieving encoded information that remains resilient to ML attacks.