It is well established that machine learning models are vulnerable to adversarial examples (AEs), which introduce imperceptible perturbations into the original image yet can significantly degrade the performance of well-trained classifiers, often resulting in nearly 100% misclassification rates. As machine learning models become more essential to society, AEs present significant safety risks to their dependability. To prevent various failure modes in real-world deployments, it is essential to develop robust defense strategies against adversarial attacks, thereby ensuring the reliability of machine learning systems.