Adaptive AI-Driven Learning Architecture for Real Time Detection and Prevention of Data Poisoning and Model Evasion Attacks
Authors:
- Paul OkandaPOPaul Okanda
- Sarah MuriithiSMSarah Muriithi
Open Access
- 4 May 2026
- journal article
- Published by UNISA Press in Southern African Journal of Security
Abstract
Artificial intelligence (AI) has been integrated into cybersecurity systems particularly in cyber threat intelligence (CTI) and has significantly enhanced the detection, analysis, and mitigation of sophisticated cyber threats in real time in organisations. Unlike traditional static defence mechanisms, which rely on fixed rules or offline retraining, the proposed architecture empowers continuous dynamic learning capabilities to identify and mitigate attacks as they unfold ensuring the integrity and reliability of AI-driven CTI systems. AI models to self-heal and harden against evolving adversarial tactics. The primary goal is to enhance CTI operations by providing timely, accurate threat detection and robust mitigation thereby reducing the window of opportunity for attackers and improving overall cybersecurity posture. This article presents a comprehensive adaptive learning architecture with formalised threat models, architectural specifications, and empirical validation. Experimental evaluation using benchmark datasets (CIC-IDS2017 and VirusTotal) demonstrates that the system achieves 43.7% higher evasion detection accuracy than traditional static anomaly detectors, while maintaining processing latency below 5 milliseconds. The architecture successfully detects 89.1% of poisoning attacks and reduces false positive rates by 37.5% relative to baseline methods. The key contributions of this article include the design of an integrated adaptive learning system that simultaneously detects and prevents both data poisoning and model evasion attacks within cybersecurity AI models.Keywords
This publication has 41 references indexed in Scilit:
- CNN-DST-IDS: CNN and D-S Evidence Theory Based Intrusion Detection System2025
- Explanatory and predictive modeling of cybersecurity behaviors using protection motivation theoryComputers & Security, 2024
- Security strategies for AI systems in Industry 4.0Quality and Reliability Engineering International, 2024
- Adversarial Training of Logistic Regression Classifiers for Weather Prediction Against Poison and Evasion Attacks2024
- Analysis of statistical properties of variables in log data for advanced anomaly detection in cyber securityComputers & Security, 2023
- Data poisoning attacks against machine learning algorithmsExpert Systems with Applications, 2022
- Exploring generative adversarial networks and adversarial trainingInternational Journal of Cognitive Computing in Engineering, 2022
- Threats on Machine Learning Technique by Data Poisoning Attack: A Survey2021
- Attack Detection Method based on Bayesian Hypothesis Testing Principle in CPSProcedia Computer Science, 2021
- Adversarial Attacks and Defenses in Images, Graphs and Text: A ReviewInternational Journal of Automation and Computing, 2020