Opto-electronic oscillators for secure encryption and efficient von-Neumann computing

We investigate the use of opto-electronic oscillators for application in encryption, the computation of difficult optimization and the training of neural networks. In encryption,we address the key distribution problem, which is the challenge of generating and sharing long encryption keys between sender and receiver in a secure manner. We develop a key distribution method based on chaotic synchronized system, in which the output of two identical chaotic systems is synchronized by a common external drive signal and then used to generate random encryption keys. We design such chaos-based key distribution based on opto-electronic oscillators and demonstrate that this allows to generate and shareone-time-pads at Gbit/s rates and low bit-error-ratios. Compared to state-of-the-art laser-based methods, the opto-electronic system provides significant improvements as it is highly stable and does not suffer from restrictions in the key generation rate due to relaxation oscillations, which can provide an up to five fold increase in the key generation rate. We test the security of the system and find that no information about the generated key can be distilled from the public channels by an attacker. Furthermore, we find synchronizationto be highly sensitive to changes in the physical parameters of the systems. We provide a general method to arbitrarily enhance the number of physical parameters, which makesstealing of the encryption key by an attacker highly unlikely. Based on an opto-electronic feedback system, we also design and build an analog Ising machine for solving difficult optimization problems. Such Ising machines work based on the insight that various NP-hard optimization problems can be mapped to a simple spin system and then solved by the natural tendency of physical spin system to evolve to their lowest energy configuration. By driving the OEO into a bistable operating regime, we are able to implement spin states in the amplitude of the photovoltage. We develop an FPGA-based time-multiplexing scheme,which allows us to couple several OEOs to implement arbitrary Ising spin networks. We find that our Ising machine achieves equal and even improved performance in NP-hardoptimization problems compared to other state-of-the-art systems. Compared to these systems, the opto-electronic Ising machine is significantly compacter and inexpensive to build and can be miniaturization with photonic integration technology, which promises a leap in the practicability of Ising machines. We also show that the nonlinear system used in our Ising machine provides an inherent advantage in computational performance, with order-of-magnitude improvements in the time-to-solution. Furthermore, we extend the range of applications for Ising machine by developing a method for their efficient use in
statistical sampling tasks through the injection of noise. By using analog Ising machines,potential order-of-magnitudes in speedup can then be achieved over conventional Monte Carlo sampling methods. We apply this sampling to the training of neural networks and find that Ising machines can achieve equal performance to software-based sampling techniques.