Detection and prevention of evasion attacks on machine learning models

With the increasing use of machine learning models in critical applications such as image classification, natural language processing, and cybersecurity, there is a growing concern about the vulnerability of these models to adversarial attacks. Evasion attacks, in particular, pose a significant threat by manipulating input data to mislead the model's predictions. This paper presents an overview of evasion attacks on machine learning models, its variants and conducts an adaptive white-box evasion attack to highlight how defense measures can be superseded with stronger evasion attack algorithms. It first discusses the different types of evasion attack algorithms, including Fast Gradient Sign Method (FGSM), Projected Gradient Descent (PGD), and Carlini–Wagner, highlighting their impact on model performance and security. It then reviews various detection and mitigation techniques which aim to identify adversarial examples and improve model robustness. Finally, the paper proposes effective mitigation techniques against evasion attacks and recommends a machine learning based cybersecurity architecture workflow that can be practically applied by organizations in real-world settings. Overall, this paper provides a comprehensive overview of evasion attacks on machine learning models and highlights the current state of research in defending against them.