A SMART ZERO TRUST SECURITY FRAMEWORK FOR COMBATING AI-DRIVEN CYBERATTACKS IN FINANCIAL INSTITUTIONS

Cybercriminals are currently weaponising Artificial Intelligence (AI) to execute convoluted cyberattacks. This new type of cyberattack is known as an AI-driven attack. AI-driven attack incorporates AI into conventional cyberattack tools to elude detection and inflict more damage. Few studies have demonstrated the effectiveness of zero trust security frameworks and AI approaches in combating sophisticated cyberattacks. However, the existing approaches are prone to data poisoning, model weight attack, and data leakage. This study proposed a Smart zero trust security framework for combating AI-driven attacks in financial institutions to address the gaps in the existing approaches. To achieve this, the study investigated the Central Bank of Nigeria risk-based cybersecurity framework to examine the use-case, stakeholders responsibilities, and reusable concepts. The study designed a DevOpsSec technique to distribute security across the development phase. A systolic addressing approach was implemented to ensure continuous threat hunting. The Federated Artificial Intelligence Technology Enabler Framework was adopted to create virtual banks and a central server. The virtual banks collaborate to train the model under the supervision of the central server without exposing their data to others. The Gradient Boosting Decision Tree and SecureBoost techniques were used to train the model. At the same time, the model-agnostic post-hoc explainer was used to explain essential features that influence the model decision. The proposed model was trained on the Zeek and Intelligent Security Group Dataset and the Nigerian Banks dataset. The systolic addressing was simulated in a network lab environment. The implemented model was evaluated with standard machine learning evaluation metrics and benchmarked with state-of-art approaches. The result shows that the implemented model achieved the best performance, with 99.81% and 99.99% prediction accuracy, 100% precision, recall and F1-score for the binary classification on the Zeek and Intelligent Security Group Dataset and Nigerian Banks Dataset. The systolic addressing was able to detect malicious patterns in 56.14 seconds. The model agnostic post-hoc explainer reveals that the “flow_duration_milliseconds” positively impacts detecting AI-driven attacks, while the packet sent has a decreasing effect. The model was also evaluated with the ISO/IEC 27000:2018 cybersecurity vulnerability assessment techniques such as the Common Vulnerability Scoring System and DDoS Resiliency Score. The model achieved a Common Vulnerability Score of 3.15 and a DDoS Resiliency Score of 7.0. This implies that the model is capable of withstanding multiple variant attacks. The result suggests that the model can efficiently be incorporated into the existing zero trust security policy engine to enhance protection