What is it about?
Audio adversarial examples appear indistinguishable from benign audio waves so that they can deceive automatic speech recognition (ASR) systems to decode them as intentional malicious commands. This paper proposes a novel method for detecting audio adversarial examples by adding noise to the logit.
Featured Image
Photo by Michael Maasen on Unsplash
Why is it important?
Audio adversarial examples become more robust to be used in over-air attacks. Most previous researches have either failed to handle various types of attacks effectively or have resulted in significant time overhead. Our approach introduces noise to the logit based on the finding that the logit of adversarial examples has a weakness to noise. We feed both smoothed audio and original audio inputs into the ASR system, and compare the decoded results whether they show an acceptable similarity.
Perspectives
I believe this article has shown a new novel approach in making robust AI system. Most approaches focused on building robust AI engine by considering relationships between neural networks while training. Our approach focuses on manipulating the last logit layer while recognizing inputs.
Jong Kim
Read the Original
This page is a summary of: Toward Robust ASR System against Audio Adversarial Examples using Agitated Logit, ACM Transactions on Privacy and Security, April 2024, ACM (Association for Computing Machinery),
DOI: 10.1145/3661822.
You can read the full text:
Contributors
The following have contributed to this page
