Learning-based optical authentication in complex scattering media

In recent years, optical encryption has become a promising method for securing information. However, optical cryptosystems have been demonstrated to be vulnerable to the attacking algorithms. Therefore, it is highly desirable that new optical cryptosystems can be continuously developed to achieve the higher security and withstand the attacks. In this paper, learning-based optical authentication in complex scattering media is proposed and experimentally verified. The ciphertext is generated by using an optical setup in complex scattering media, and a learning model is developed and trained to generate security keys. Moreover, other parameters, e.g., virtual phase-only masks, can be flexibly designed and used to further enlarge key space. The training pairs fed to the designed leaning model consist of specially processed images (i.e., acting as plaintexts) and speckle patterns (i.e., ciphertexts) recorded by CCD. The retrieved plaintexts obtained from the trained learning model cannot visually render recognizable information, and can be effectively authenticated by using nonlinear correlation algorithm. Optical experiments are conducted to verify effectiveness and applicability of the proposed learning-based optical authentication in complex scattering media to achieve the higher security and withstand the attacks.