Closeness and Uncertainty Aware Adversarial Examples Detection in Adversarial Machine Learning

Machine learning (ML) applications are transforming our everyday lives and the artificial intelligence technology is becoming an integral part of our civilization. As the artificial intelligence technology advances, it becomes a key component of many sophisticated tasks that have direct effect on humans. In the last few years, deep neural networks (DNNs) achieved state-of-the-art performances on different number of supervised learning tasks, which led them to become widely used in many fields such as medical diagnosis, computer vision, machine translation, speech recognition and autonomous vehicles [1, 2, 3, 4]. However, there exist serious concerns on how to make deep neural networks an integral part of our lives while ensuring utmost security and reliability. Although DNNs have proven their usefulness in the real-world applications for many complex problems, they have thus far failed to overcome the challenges faced by deliberately manipulated data, which are known as adversarial inputs [5].