EFFICIENT AND TRUSTWORTHY ANTI-CORRUPTION AI: LEVERAGING MODEL PRUNING FOR SECURE DEPLOYMENT IN CRITICAL SYSTEMS

Abstract

Corruption in public procurement, financial transactions, and administrative processes presents substantial threats to national and international security, compromising integrity in defense logistics, resource distribution, and governance frameworks. Artificial Intelligence (AI), particularly machine learning models for anomaly detection, fraud identification, and risk prediction, has emerged as a powerful analytical tool for procurement risk monitoring. However, large-scale AI models face significant barriers to deployment in secure, resource-constrained environments such as edge devices used for military or public-sector monitoring. These barriers include high computational demands, susceptibility to adversarial perturbations, and limited interpretability — all of which reduce their suitability for security-critical applications.
This study examines model pruning techniques, including structured and unstructured pruning as well as magnitude-based and lottery ticket hypothesis-inspired approaches, to create compact and efficient anti-corruption AI models with minimal accuracy degradation. These pruned models achieve faster inference and substantially reduced memory footprints, enabling deployment on edge devices and in resource-constrained secure environments. Experimental results demonstrate effective identification of key procurement risk signals, including single-bid awards, suspiciously short tender windows, and irregular fund flows. The proposed approach offers a pathway for integrating compact, efficient procurement anomaly detection models into defense and security technologies, supporting decision support systems and infrastructure protection. This work bridges advances in artificial intelligence with secure technology development, contributing to enhanced societal integrity and resilience.