Adversarial attacks on machine learning-based cyber security systems: a survey of techniques and defences
Pratik Shashikant Patel, Pooja Panchal · International Journal of Electronic Security and Digital Forensics · 2024
Machine learning (ML) has been increasingly adopted in the field of cyber security to enhance the detection and prevention of cyber threats.However, recent studies have demonstrated that ML-based cyber security systems are vulnerable to adversarial attacks, in which an attacker manipulates input data to deceive the ML model and evade detection.This paper presents a survey of adversarial attacks on ML-based cyber security systems, including techniques such as evasion, poisoning, and backdoor attacks.Additionally, we discuss the limitations of current defences against adversarial attacks, such as defensive distillation and adversarial training, and propose potential future directions for defence mechanisms.Finally, we provide a framework for evaluating the effectiveness of existing defences against adversarial attacks on ML-based cyber security systems.Our survey highlights the urgent need for developing more robust and reliable defence mechanisms to ensure the security and reliability of ML-based cyber security systems in the face of adversarial attacks.