ASQ-FastBM3D: An Adaptive Denoising Framework for Defending Adversarial Attacks in Machine Learning Enabled Systems

Abstract

Machine learning has made significant progress in image recognition, natural language processing, and autonomous driving. However, the generation of adversarial examples has proved that the machine learning system is unreliable. By adding imperceptible perturbations to clean images can fool the well-trained machine learning systems. To solve this problem, we propose an adaptive image denoising framework Adaptive Scalar Quantization ( <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">ASQ-FastBM3D</i> ). The <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">ASQ-FastBM3D</i> framework combines the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">ASQ</i> method with the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">FastBM3D</i> algorithm. The adaptive scalar quantization is the improvement of scalar quantization, which is used to eliminate most of the perturbations. <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">FastBM3D</i> is proposed to improve the quality of the quantified image. The running time of <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">FastBM3D</i> is 50% less than that of <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">BM3D</i> . Compared with some traditional filter methods and some state-of-the-art neural network methods for recovering the adversarial examples, the accuracy rate of our <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">ASQ-FastBM3D</i> method is 99.73% and the F1 score is 98.01%, which is the highest.

Topics & Concepts