Reconfigurable Silicon Photonics Extreme Learning Machine with Random Non-linearities as Neural Processor and Physical Unclonable Function

TL;DR

This paper introduces a reconfigurable silicon photonics-based extreme learning machine utilizing random non-linearities for efficient machine learning and secure physical unclonable functions, demonstrating high performance and security.

Contribution

It presents a novel RN-ELM architecture using integrated silicon photonics with random non-linearities, enabling compact design, high accuracy, and secure authentication.

Findings

Achieves state-of-the-art time-series prediction with minimal hardware.

Demonstrates low clone probability (10^-15) for secure authentication.

Uses reconfigurable silicon photonic mesh for flexible non-linearity implementation.

Abstract

An alternative extreme learning machine -ELM- paradigm is presented exploiting random non-linearities -RN, named RN-ELM, instead of a conventional fixed node non-linearity. This method is implemented on a hybrid neural engine, with the physical layer realized by an integrated silicon photonic mesh and the digital layer by a simple regression algorithm. Non-linearities are intrinsically non-power depended and are generated through non-linear frequency to power mapping offered by optical filters. The numerical evaluation is based on an experimentally derived transfer function of an all-pass filter, implemented on a silicon reconfigurable photonic integrated chip -RPIC. RN-ELM is evaluated in a twofold manner; first as a machine learning scheme, where the expressivity offered by multiple, yet random, activation functions lead to a compact and highly simplified design with 5 optical filters, offering state-of-the-art performance in time-series prediction tasks with minimum hardware requirements. The second scenario entails its deployment as a physical unclonable function -PUF, for authentication applications directly in the physical layer. In this case, the random activation functions are associated with unavoidable, fabrication related waveguide imperfections that can act as hardware signatures. Numerical results reveal a probability of cloning as low as 10e-15, which corresponds to a highly secure authentication token.