Elliptic Curve Cryptography (ECC) is a technology based on the arithmetic of elliptic curves used to build strong and efficient cryptosystems and infrastructures. Several ECC systems, such as the Diffie–Hellman key exchange and the Elliptic Curve Digital Signature Algorithm, are deployed in real-life applications to enhance the security and efficiency of digital transactions. ECC has gained even more importance since the introduction of Bitcoin, the peer-to-peer electronic cash system, by Satoshi Nakamoto in 2008. In parallel, the integration of artificial intelligence, particularly machine learning, in various applications has increased the demand for robust cryptographic systems to ensure safety and security. In this paper, we present an overview of machine learning and Elliptic Curve Cryptography algorithms. We begin with a detailed review of the main ECC systems and evaluate their efficiency and security. Subsequently, we investigate potential applications of machine learning-based techniques to enhance the security and performance of ECC. This study includes the generation of optimal parameters for ECC systems using machine learning algorithms.

In today's digital age, mobile applications have become essential in connecting people from diverse domains. They play a crucial role in enabling communication, facilitating business transactions, and providing access to a range of services. Mobile communication is widespread due to its portability and ease of use, with an increasing number of mobile devices projected to reach 18.22 billion by the end of 2025. However, this convenience comes at a cost, as cybercriminals are constantly looking for ways to exploit security vulnerabilities in mobile applications. Among the several varieties of malicious applications, zero-day malware is particularly dangerous since it cannot be removed by antivirus software. To detect zero-day Android malware, this paper introduces a novel approach based on generative adversarial networks (GANs), which generates new frequencies of feature vectors from system calls. In the proposed approach, the generator is fed with a mixture of real samples and noise, and then trained to create new samples, while the discriminator model aims to classify these samples as either real or fake. We assess the performance of our model through different measures, including loss functions, the Frechet Inception distance, and the inception score evaluation metrics.

Generative adversarial networks (GANs) are a technology that uses two neural networks against each other in a maximization game to generate new instances of data that look similar to actual data. The GANs was first proposed by Ian Goodfellow et al.[10]. From there, researchers have presented many models of GANs, such as Wasserstein GAN, conditional GAN, deep convolutional GAN, and many others models. Every model has its character and architecture, and there are some elements in common. One neural network is called the generator; it generates new data samples that have the same probability distribution as the distribution of training data. While the other is called the discriminator, its role is to determine whether the input data came from the training dataset or were created by the generator. In other words, it classifies if the generated data that was caused by the generator is real or fake. The GANs technology was used widely in image and video generation, but lately, this technique has broken into the mobile security world. Especially for android, since it is an open-source operation system, it provides hackers with considerable freedom to develop new types of attacks and malware. In this paper, we provide a comparison between the well-known GANs models: Wasserstein GAN, conditional GAN, and deep convolutional GAN. We explain their architectures, objectives functions, and critical situations in which they can be used for the security of mobile applications

Malware detection is an escalating challenge, prompting researchers to propose numerous solutions leveraging various techniques as crucial components in combating malware. Intrusion System Detection (IDS) assumes a central role in fortifying security and is deemed an essential defense strategy for every infrastructure. However, regardless of the IDS method employed, novel challenges persist, including the imperative for minimal CPU time, indicative of the IDS response time. In this paper, we present an optimal Black Box IDS based on machine learning classifiers and obfuscated features. This streamlined approach incorporates an innovative algorithm that reduces the feature set using the Chi-Square distribution, achieving maximum accuracy and minimal CPU time simultaneously. The proposed IDS implementation involves four machine-learning classifiers, namely Naïve Bayes, Support Vector Machines, K-Nearest Neighbors, and Random Forest. Our tests demonstrated the efficiency of the proposed approach, effectively minimizing features to a crucial set of five, attaining an impressive accuracy rate of 99.90% within a CPU time corresponding to 0.0017 seconds. This outcome underscores the efficiency of our approach in significantly reducing the feature set while maintaining high accuracy, highlighting its potential for effective implementation in practical scenarios.