Demystifying the Black Box: A Foundational Survey and Comparative Analysis of Basic Explainable AI (XAI) Techniques

作者：Veena More, Bharati H Naikawadi, Kella Sowmya, Sumitra M Mudda · 发表于：International Journal of Innovative Research in Computer and Communication Engineering · 年份：2025 · DOI：10.15680/ijircce.2025.1312094

ABSTRACT: The unprecedented performance of complex Artificial Intelligence (AI) models, particularly deep neural networks (DNNs), has been accompanied by a critical opacity problem—their decision-making processes are often inscrutable "black boxes." This lack of transparency erodes trust, impedes debugging, and prevents deployment in highstakes domains like healthcare, finance, and criminal justice, where accountability is paramount. Explainable AI (XAI) has emerged as a vital subfield dedicated to making AI systems more interpretable and understandable to human stakeholders. This research paper provides a comprehensive, pedagogical exploration of fundamental, post-hoc XAI techniques, designed to demystify the core concepts for researchers and practitioners entering the field. We systematically categorize and describe essential methods spanning feature importance techniques (e.g., Permutation Feature Importance, SHAP, LIME), example-based explanations (e.g., Counterfactual Explanations, Prototypes/Criticisms), and visualization techniques for deep learning (e.g., Saliency Maps, Grad-CAM). The study employs a structured methodology of theoretical exposition followed by a consistent empirical evaluation across three benchmark datasets (tabular, image, text) using three common model types (Random Forest, Convolutional Neural Network, BERT). We implemented and compared eight basic XAI techniques, analyzing their outputs not just on predictive accuracy but on key explainability cr...