| Abstract: | Audio event detection systems capture audio from the environment and detect the presence of a specific sound of interest. These systems make extensive use of deep learning classifiers as their primary detection algorithm. We first developed deep learning-based models for detecting physical safety-related audio events in this project, such as gunshot, glass breaking, car alarm, and dog barking. Secondly, we evaluated their robustness against adversarial examples, where the adversary tries to avoid detection, and finally, we proposed new methods for increasing the robustness of these systems against adversarial examples. |