Toward Secure and Efficient CNN Recognition with Different Activation and Optimization Functions

Abstract

Precise data input classification and identification are essential for developing high-quality intelligent systems, including defending systems against cyberattacks and guaranteeing data protection. The Internet of Things (IoT), in particular, offers specific problems for cybersecurity not because of the volumes of data generated in a short time but also for the diversity of data sources. Thus, creating an accurate identification system is crucial to preserve data integrity and security. It has been demonstrated that convolutional neural networks (CNNs) perform well in developing data-driven applications, including cybersecurity. However, hyperparameters play a significant role in their prediction quality which may increase/decrease the false alarm rates. CNNs’ performance may be improved by the activation and optimization functions utilized using nonlinear expression and reducing loss functions, enabling them to be more precisely fit data. This study evaluates the classification accuracy of CNNs while classifying inputs using various combinations of activation functions and optimization strategies. The performance of the CNN model with ReLU and PReLU activation functions that utilize RMSprop and Adam optimization strategies is examined. However, even the most minor enhancement may significantly impact cybersecurity. A handwriting text is one of the authentication mechanisms to identify persons by their signature and writing style. In this experiment, a CNN model with Adam optimizer and PReLU activation function achieved the highest accuracy of 98.60% when trained on the Kaggle handwritten digit dataset, making a significant difference compared to classical learning algorithms and a slight improvement to other CNN models.