Constructing adversarial samples against deep learning-based sensing system (part II)

• Main Author: Lin, Beiyi
• Other Authors: Tan Rui
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2020
• Subjects: Engineering::Computer science and engineering
• Online Access: https://hdl.handle.net/10356/137412
• Institution: Nanyang Technological University

As deep learning become more popular and have grown to become crucial components in the daily device we use. Despite their effectiveness, they are not invincible. Adversarial examples initially discovered and applied to computer vision systems are now becoming a noticeable issue in DeepSpeech processing classifier as well. Adversarial examples are input samples that have very bad classification accuracy. They are generated by adding imperceptible perturbations by human. These adversarial examples result in misclassification of results. In late 2017, an attack was shown to be quite effective against the Speech Commands classification model. Speech commands are used very frequently in many applications, such as Google Assistant, Amazon Alexa and Apple's Siri. Thus, adversarial examples produced by this attack could have real-world consequences. While previous work in defending again these malicious attacks has investigated using gradient masking to hide information of the model and audio pre-processing to reduce or distort adversarial noise, this project explores the idea of simple pink noise injection at different loudness to detect adversarial examples. This technique of noise injection does not require retraining or modifying the model. It is also possible to transfer this technique from the model used in this project to other DeepSpeech models.