V2I8P16

Detecting and Defending Against Adversarial Attacks on AI Powered Security Systems

Saeed Hubairik Aliyu^1*, Abosede Oluwabunmi Adelore², Chijioke C. Ekechi³, Hanafi Musa Olayinka⁴, Oluwasola Abiodun Elijah⁵, Unuigbokhai Peter Afegbai⁶

Abstract

Adversarial attacks pose a critical threat to the security and reliability of artificial intelligence (AI) powered systems, exploiting model vulnerabilities to manipulate outcomes and compromise sensitive operations. These attacks (ranging from evasion and data poisoning to model extraction and prompt injection) undermine the integrity of AI models across high stakes applications, including healthcare, autonomous systems, and cybersecurity. This paper explores the taxonomy of adversarial attacks, their operational impact, and the current landscape of defence strategies designed to detect and mitigate them. Emphasis is placed on proactive and adaptive approaches such as adversarial training, real time monitoring, input preprocessing, and explainable AI techniques. While notable progress has been made, the evolving nature of these threats continues to challenge existing defences, demanding more scalable, interpretable, and resource efficient solutions. The study underscores the urgency of cross disciplinary collaboration to develop resilient AI systems capable of withstanding adversarial manipulation in increasingly complex environments.

Keywords:

Adversarial Attacks, AI Security Evasion Attacks, Data Poisoning, Model Extraction, Prompt Injection, Intrusion Detection, Adversarial Training, Input Preprocessing, Explainable AI (XAI), Machine Learning, Robustness, AI Threat Mitigation, Secure AI Systems