<![CDATA[In the realm of Internet of Things (IoT) systems, the generation of cryptographic keys is crucial for ensuring secure data transmission and device authentication. However, traditional methods of generating random keys encounter challenges about security, efficiency, and scalability, particularly when applied to resource-constrained IoT devices. To address these issues, neural networks have emerged as a promising approach due to their ability to learn intricate patterns. Nonetheless, the architecture of neural networks significantly impacts their performance. This paper presents a comprehensive comparative analysis of three commonly employed neural network architectures for generating cryptographic keys on IoT devices. We propose a novel neural network-based algorithm for key generation and implement it using each architecture. The models are trained to generate cryptographic keys of various sizes from random input data. Performance evaluation is conducted based on key metrics such as accuracy, loss, key randomness, and model complexity. Experimental results indicate that the Feedforward Neural Network (FFNN) architecture achieves exceptional accuracy of over 99% and successfully passes all randomness tests, surpassing the alternatives. Convolutional Neural Networks (CNNs) demonstrate subpar performance as they emphasize spatial features that are irrelevant to key generation. Recurrent Neural Networks (RNNs) struggle with the complex long-range dependencies inherent in generating keys]]>
