The field of machine learning has seen a growing concern regarding the security of deep learning models. Adversarial attacks pose a significant threat to the integrity and reliability of these models. A Large number of these attacks are based on gradient-based approaches. Many defensive mechanisms have been proposed to counter gradient-based attacks. Dynamic defenses are the type of defenses against gradient-based adversarial attacks in which the parameters of the deep learning models are updated during the prediction to make the model’s gradient different than the original one, thus, misleading the adversary. Dent (Dynamic entropy minimization) and Unig (Unifying gradient) are two examples of dynamic defenses. In the first part of this project, we implement the Unig defense technique and perform experiment using Unig and Dent on a ResNet18 model trained on Cifar10 dataset. In the second part of the research, we propose two new dynamic defense mechanism with gradient distortion of custom added layers. The experimental results using the proposed dynamic defenses have also been provided.