AI vs AI: Using artificial intelligence to protect data and attacks on it

Abstract

The aim of the study was to conduct a comprehensive analysis of the role of artificial intelligence as a tool for attacks and defence in information systems, with a focus on evaluating the effectiveness of existing approaches and justifying prospects for integrating artificial intelligence to enhance cybersecurity under conditions of increasing intelligent threats. Within the framework of the study, the confrontation between offensive and defensive artificial intelligence systems in a dynamic environment with adaptive behaviour was modelled, which made it possible not only to identify typical threat vectors but also to assess the effectiveness of corresponding countermeasures. It was found that generative models, particularly those based on reinforcement learning, effectively adapted to defensive responses, bypassing traditional filters and heuristics. At the same time, the highest resilience to such attacks was demonstrated by combined approaches that integrated Federated Learning, blockchain, and differential privacy: the level of resistance to attacks increased by up to 40% with a moderate decrease in accuracy (3-6%). Adversarial training ensured an increase in security of up to 25%, although the accuracy dropped by up to 4%, and its effectiveness depended significantly on the completeness and variability of training data. Homomorphic encryption proved to be the most confidential approach, but remained limited in practical use due to excessive resource consumption and processing time. Although blockchain tools contributed to transparency and data immutability, these tools involved high latency, which complicated the application in real-time conditions. Overall, the results of the study confirmed the appropriateness of using multimodal, adaptive, and multi-level protection strategies for artificial intelligence systems, especially amid the growing number of generative attacks, as evidenced by real cases (e.g., Disney). The practical significance lay in the formation of foundations for the development of adaptive cyber defence systems capable of countering intelligent attacks in real time. The obtained results could be used to enhance the security of critical infrastructure, financial platforms, and autonomous systems.