Mitigating Adversarial and Data Poisoning Attacks in Cybersecurity Applications
Saad Rustam, Usama Arshad, Muhammad Zaini Ahmad, Najam Sawera, Hammad Ali · 2024
The integration of deep learning models in cybersecurity applications, such as malware detection, intrusion detection, and spam filtering, has introduced significant vulner-abilities to security threats. Among these, adversarial and data poisoning attacks pose particularly severe risks, compromising the reliability and accuracy of learning algorithms. Adversarial attacks involve the manipulation of inputs to deceive models into making incorrect predictions, while data poisoning attacks aim to corrupt training datasets, thereby degrading model performance. This paper provides an in-depth analysis of these vulnerabilities, offering a comprehensive explanation of the different types of adversarial and data poisoning attacks. We also explore existing defense mechanisms, including effective model training techniques, anomaly detection, and secure data processing. Our study critically evaluates the practical efficacy of these defenses against common cyber threats, highlighting both their strengths and limitations. By addressing these critical challenges, our research contributes to enhancing the security and robustness of machine learning models in combating today's sophisticated cyber threats. The findings and recommendations presented in this paper will be of significant interest to researchers and practitioners striving to bolster machine learning's resilience against adversarial attacks and ensure data authenticity.