The paper addresses the need for faster and more efficient decryption methods in modern communication systems. Current electronic decryption processes bottleneck the speed of optical communication. The authors propose using optical machine learning to perform decryption directly in the optical domain, leveraging the inherent advantages of speed and parallelism. Traditional optical security schemes using phase masks are limited by their rigidity and bulkiness. The integration of machine learning offers a path towards compact, flexible, and high-speed decryption systems. The proposed optical machine learning decryptors (MLDs) are trained to recognize specific keys (symmetric decryption) or classes of keys (asymmetric decryption) for enhanced authentication. These MLDs are designed as single-layer holographic perceptrons, physically fabricated using galvo-dithered two-photon nanolithography (GD-TPN) for high neuron density and on-chip integration with CMOS.

The introduction reviews existing optical security schemes based on phase masks and their limitations, highlighting the need for more flexible and compact systems. It also discusses prior work on optical implementation of matrix multiplication using various methods such as beam splitters, Mach-Zehnder interferometers, and integrated photonic circuits. The limitations of existing diffractive neural network architectures with low neuron density are addressed, motivating the use of nanolithography for creating more compact devices. The advantages of GD-TPN over other nanolithographic methods and its suitability for integrating with CMOS are emphasized.

The methodology details the design and training of the MLDs using a computer-based machine learning approach. The MLD is modeled as a three-layer diffractive neural network (input, MLD, output), with neurons linked through Rayleigh-Sommerfeld diffraction. The training process involves optimizing the phase delay of each neuron in the diffractive layer to minimize a cross-entropy loss function using a TensorFlow-based approach. The paper explores both single-layer MLDs and compact multilayer perceptrons (CMPLs), comparing their performance. Symmetric and asymmetric decryption schemes are implemented and tested. The fabrication process uses GD-TPN to nanoprint the MLDs onto CMOS chips, achieving high neuron density and precise control over neuron height. The experimental setup for evaluating the MLDs' performance uses a 785 nm laser source, spatial light modulator (SLM), and charge-coupled device (CCD) camera.

The key findings include the successful fabrication of MLDs with a neuron density exceeding 500 million neurons per square centimeter using GD-TPN. Experimental results demonstrate the ability of the nanoprinted MLDs to perform both symmetric and asymmetric decryption with high accuracy. Symmetric decryption is demonstrated using two decryptors (MLD-T and MLD-B) achieving 100% accuracy. Asymmetric decryption using 3-MLD and 9-MLD shows experimental accuracies of 86.67% and 80%, respectively. The study shows that the number of neurons, neuron density, and distances between layers significantly impact the MLD's performance. The analysis includes detailed characterization of the nanoprinted MLDs using atomic force microscopy (AFM). The diffraction efficiency and accuracy were calculated with varying degrees of noise added to the experimental data.

The results demonstrate the feasibility of integrating high-performance optical machine learning decryptors directly onto CMOS chips. The achieved neuron density significantly surpasses existing diffractive neural networks, highlighting the potential for more complex and powerful optical computing systems. The successful implementation of both symmetric and asymmetric decryption showcases the versatility of the MLD architecture for various security applications. The close match between experimental and numerical results validates the accuracy of the computational models and the effectiveness of the nanoprinting technique. Future work could explore more complex network architectures and applications in other fields such as optical sensing and medical diagnostics.

This research successfully demonstrates the creation and implementation of high-neuron-density optical linear perceptrons for near-infrared inference tasks on a CMOS chip. The use of GD-TPN enabled the fabrication of devices with unprecedented neuron density, demonstrating significant advancements in optical machine learning. Future research could investigate more complex network architectures and expand the range of applications beyond decryption, potentially transforming various fields.

While the study demonstrates high accuracy for some decryption tasks, the accuracy for asymmetric decryption with a larger number of classes is lower. The influence of noise on the system's performance, although investigated, could be further studied. The current study focuses on specific types of images (handwritten letters); the generalizability to other types of input data should be further explored.