INTEGRATION OF EXPLICABLE AI (XAI) INTO THE PROTECTED CONTROL CIRCUITS OF THE SYSTEMS

Abstract

The integration of machine learning algorithms and deep neural networks into safety-critical cyber-physical systems (CPS) requires a transition from stochastic “black-box” models to transparent architectural solutions that ensure deterministic safety. The technical necessity of integrating explainable artificial intelligence (XAI) is driven by verification and validation (V&V) requirements in non-deterministic environments where the cost of error precludes the use of unverified heuristics.

This work investigates the fundamental conflict between the predictive power of models and their real-time interpretability. Taxonomies of XAI are examined, including intrinsic and post-hoc methods, as well as their impact on latency in control loops. Architectural trade-offs associated with the use of hardware accelerators based on FPGA and specialized ASICs are analyzed.

Particular attention is given to the robustness of explanation modules against adversarial influences such as fairwashing and feature manipulation. Approaches to developing adaptive explanation systems based on implicit operator feedback are synthesized. The paper concludes by identifying critical technological gaps in the standardization of interpretability metrics and in ensuring the security of XAI algorithms themselves .