Photonic Extreme Learning Machine‐Based Neural Cryptosystem for Error‐Free Secure Data Delivery

Abstract

ABSTRACT Conventional optical cryptosystems rely on computationally heavy, detached deep‐learning black boxes and visual metrics, fundamentally hindering precise information delivery and exposing latent vulnerabilities. To overcome this, we propose a hardware‐algorithm isomorphic photonic neural cryptosystem tailored for exact data delivery. This architecture intrinsically utilizes physical optical scattering as the random projection of an Extreme Learning Machine (ELM). By synergizing the ELM's rapid closed‐form analytical inference with the deterministic error tolerance of structured data encoding, we successfully bridge the gap between statistical neural network inference and cryptographic correctness, achieving a 100% Decryption Success Rate (DSR). Furthermore, multi‐synaptic connections are repurposed as an additional physical key dimension, enlarging the key space without incurring digital computational overhead. Generating high‐entropy ciphertexts (>7 bits) and massive‐length unclonable physical keys (>96.9 megabits), the hardware‐rooted Physical Unclonable Function (PUF) coupled with a digital probabilistic encoding (salting) strategy neutralizes advanced cryptanalysis, including chosen‐plaintext attacks (CPA), known‐plaintext attacks (KPA), and machine‐learning model‐extraction attempts. Requiring only ∼0.76 M parameters and 0.67 ms inference latency, this lightweight framework natively supports high‐frequency dynamic key updates and poses a substantial computational barrier (digital keys >6 megabits) to brute‐force attacks, offering a scalable and highly efficient solution for secure communications in edge computing and IoT.