Abstract
Based on the opposition with small input perturbation, with such well crafted attacks, predictions may often be steered strongly in the wrong direction and even erratic; therefore, machine learning models are highly susceptible to adversarial attacks. Some of these techniques have trade-offs, such as less model accuracy on clean data, increased computational cost or vulnerability to novel attack techniques. It might not however be possible to implement a strong and effective countermeasure to such attacks in the real world application. To solve the above-mentioned issues, the following research aims at creating a strong adversarial defense mechanism by combining various techniques such as adversarial training, feature smoothing, and ensemble learning. A proposed technique was a combination of these techniques to boost the resilience of the models to all the varieties of adversarial attacks with the quality maintained in terms of accuracy and computational efficiency. Develop defense mechanisms to make machine learning models more resilient to adversarial attacks without making a substantial impact on them.