Architectures of secure university infrastructure for the implementation of artificial intelligence considering data confidentiality and cybersecurity requirements

Abstract

The article examines architectural principles for building secure university digital infrastructure when implementing artificial intelligence systems under heightened requirements for confidentiality and cyber resilience. It is shown that the transformation of a university into a platform ecosystem increases the risks of compromising the digital footprint of students and staff, which includes not only academic performance but also behavioral, biometric, and psychometric data, and exacerbates the contradiction between the openness of the academic environment, user mobility, and the need for strict access control. The analytical part relies on a systematic comparison of practices from 45 universities (2019-2023) using threat modeling and risk ranking, vulnerability auditing, stress testing, and metrics for detection and response timeliness; scenarios of attacks on machine learning are considered separately, including data poisoning and adversarial impacts. Based on comparative assessments, compromises between performance and protection are demonstrated for computations placed on the periphery, in a local data processing center, and in the cloud: using public cloud without secure channels shows moderate delays (45.3 ms) with elevated probability of successful attack (12.7%), whereas incorporating VPN and deep inspection reduces risk (4.2%) but leads to a sharp increase in delays (189.6 ms) and gateway load; the local circuit provides the best balance (24.1 ms; 0.8%), while physically isolated segments minimize external threats (0.02%) at the cost of functional limitations. A shift in budget priorities from hardware protection to process measures is identified: increased investments in staff training and AI model protection yield the highest returns (ROI 4.89 and 3.41), consistent with data on the high effectiveness of AI-generated phishing among bachelor's students (21,3% credential compromise rate at 4,2% incident reports). The discussion of privacy-preserving technologies quantitatively demonstrates the cost of privacy: differential privacy and federated learning incur manageable losses, whereas homomorphic encryption and multi-party computation radically increase resource intensity (up to 145-fold increase in time and hundreds of gigabytes of memory). As a practical conclusion, the preference for hybrid architectures based on Zero Trust, data and model protection throughout the lifecycle, and integration of technical mechanisms with organizational practices is substantiated, since isolated perimeter strengthening yields diminishing returns amid the evolution of AI-enhanced attacks.