This paper provides a comprehensive review of gradient-based feature attribution methods in Explainable AI (XAI). The authors highlight the growing need for explainability in black-box AI models, particularly in high-stakes applications like healthcare and autonomous driving. They focus on gradient-based explanations, which are directly applicable to neural networks. The paper introduces a novel taxonomy to categorize these methods into four groups: vanilla gradients-based explanations, integrated gradients-based explanations, bias gradients-based explanations, and post-processing for denoising. The authors systematically explore the evolution of gradient-based explanation methods, emphasizing their technical details and algorithmic properties. They discuss both human and quantitative evaluations to measure the performance of these methods. The paper also identifies general and specific challenges in XAI and gradient-based explanations, aiming to help researchers understand the state-of-the-art progress and its limitations. The paper reviews various gradient-based explanation methods, including backpropagation, deconvolutional networks, guided backpropagation, RectGrad, and Grad-CAM. It also discusses integrated gradients, blur integrated gradients, expected gradients, split integrated gradients, integrated Hessians, integrated directional gradients, guided integrated gradients, adversarial gradient integration, boundary-based integrated gradients, and important direction gradient integration. These methods aim to explain model predictions by analyzing the influence of input features on the output. The paper also addresses the role of bias terms in shaping model outputs and introduces FullGrad, a method that decomposes the output prediction into input gradients and bias gradients. Finally, the paper discusses post-processing techniques like SmoothGrad and VarGrad to denoise explanations and improve their quality. The authors conclude that gradient-based explanations are essential for making AI models more interpretable and trustworthy, and that further research is needed to address the challenges and limitations of these methods.This paper provides a comprehensive review of gradient-based feature attribution methods in Explainable AI (XAI). The authors highlight the growing need for explainability in black-box AI models, particularly in high-stakes applications like healthcare and autonomous driving. They focus on gradient-based explanations, which are directly applicable to neural networks. The paper introduces a novel taxonomy to categorize these methods into four groups: vanilla gradients-based explanations, integrated gradients-based explanations, bias gradients-based explanations, and post-processing for denoising. The authors systematically explore the evolution of gradient-based explanation methods, emphasizing their technical details and algorithmic properties. They discuss both human and quantitative evaluations to measure the performance of these methods. The paper also identifies general and specific challenges in XAI and gradient-based explanations, aiming to help researchers understand the state-of-the-art progress and its limitations. The paper reviews various gradient-based explanation methods, including backpropagation, deconvolutional networks, guided backpropagation, RectGrad, and Grad-CAM. It also discusses integrated gradients, blur integrated gradients, expected gradients, split integrated gradients, integrated Hessians, integrated directional gradients, guided integrated gradients, adversarial gradient integration, boundary-based integrated gradients, and important direction gradient integration. These methods aim to explain model predictions by analyzing the influence of input features on the output. The paper also addresses the role of bias terms in shaping model outputs and introduces FullGrad, a method that decomposes the output prediction into input gradients and bias gradients. Finally, the paper discusses post-processing techniques like SmoothGrad and VarGrad to denoise explanations and improve their quality. The authors conclude that gradient-based explanations are essential for making AI models more interpretable and trustworthy, and that further research is needed to address the challenges and limitations of these methods.