Adversarial Attacks on AI-driven Cybersecurity Systems: A Taxonomy and Defense Strategies
Krishna Chaganti
Abstract
This article is focused on the analysis of the increasing risks of hostile cyber operations targeting AI and their relevance to cybersecurity. The study proposes a novel classifying system of adversarial ML attacks based on their aims, the knowledge level of the attacker, the degree of specialization concerning the attack target, and the frequency of the attack. It shows that the same kind of attacks are possible during both the training and inference phases and explains how small changes to data can affect ML systems. Examples like healthcare model theft and data poisoning during the recent U.S. elections highlight the serious threats to public confidence and data integrity. Nonetheless, it also suggests a strong defense framework that incorporates defense architecture with the new ideas of adversarial training, game-theoretic models, robust learning, and adversarial detection. The effectiveness of adversarial training was demonstrated by the notable 75% performance retention it attained against gradient-based assaults. To improve the system's reliability and protect it from adversarial threats, the research employs the findings of AI and cybersecurity domains. It is highly valuable for the progress of defense techniques for AI in life-critical applications and provides essential contributions to the protection of essential structures.