More and more apps are using deep learning models, which has raised worries about how vulnerable they are to threats from other programs. As a result, academics have looked into adaptable noise input methods as a way to make these models more resistant to attacks like these. In this study, we look at all the latest adaptive noise input methods that are designed to make deep learning models work better in hostile environments. By adding noise to the input space, hidden layers, or gradients during model training, these methods try to lessen the effect of hostile changes. These methods make the model more resistant to hostile manipulation without affecting its performance on clean data. They do this by changing the noise parameters on the fly based on the features of the input data or the model’s performance. This paper uses experiments and comparisons to show how well and how many different ways adaptive noise input techniques can be used to make deep learning models safer and more resilient against threats.