Abstract

Machine learning models, particularly Convolutional Neural Networks (CNNs), dominate image classification tasks in critical domains such as medical imaging, autonomous driving, and insurance. However, despite high accuracy on clean benchmark datasets, these models often exhibit significant performance degradation under real-world corruptions like noise, blur, occlusion, or compression artifacts, leading to safety risks and operational failures. Existing robustness evaluations remain limited, focusing predominantly on deep neural networks, using narrow accuracy-based metrics, and overlooking classical machine learning approaches, uncertainty quantification, prediction stability, and computational efficiency.

This thesis presents a comprehensive evaluation of seven model families—ranging from classical (Logistic Regression, SVM, K-NN, Random Forest, MLP) to deep learning (Lenet-5, ResNet-18)—on MNIST and Fashion-MNIST. We propose a unified, multi-metric framework assessing accuracy, robustness (flip rate, label variation), uncertainty (Gini index, max probability), and efficiency (parameter count, training time) under clean, corrupted, and mixed-noise conditions.

Our findings offer practical insights into model reliability and highlight the trade-offs between performance, stability, and computational cost—supporting more informed choices in real-world deployments.