深度卷积神经网络对Perlin噪声的敏感性研究  

Sensitivity of Deep Convolution Neural Network to Perlin Noise

作　　者：刘思婷 葛万成[1] LIU Siting;GE Wancheng(Tongji University,Shanghai 201800,China)

机构地区：[1]同济大学,上海201800

出　　处：《通信技术》2021年第3期672-678,共7页Communications Technology

摘　　要：自动驾驶技术的安全性和可靠性问题一直是当下研究的热点,也是自动驾驶汽车实现市场普遍性应用前必须要克服的一大难题。针对当前自动驾驶领域面临的这个困境,提出一种基于Perlin噪声的攻击方法。此方法利用Perlin噪声生成自然纹理噪声来模拟自然场景下可能遇到的扰动,使神经网络的分类器发生误判,达到降低神经网络检测精度的效果。通过在3种不同结构的神经网络结构上进行测试,发现不同结构的神经网络对Perlin噪声的敏感程度不同,且量化分析噪声参数对网络稳定性的影响效果,同时证明了所提方法的可行性和泛化性,对以后研究神经网络的性能具有一定的参考意义。The safety and reliability of automatic driving technology has always been the focus of current research.It is also a major problem that must be overcome before realizing the universal application of autopilot.In view of the current difficulties in the field of automatic driving,this paper proposes an attack method based on Perlin noise.This method uses Perlin noise to generate natural texture noise to simulate the possible disturbance in the natural scene,so as to make the neural network classifier misjudge and achieve the effect of reducing the detection accuracy of neural network.In this paper,three different neural network structures are tested to detect the sensitivity of different neural networks to Perlin noise,and the influence of noise parameters on the stability of the network is analyzed quantitatively.At the same time,the feasibility and generalization of this method are proved.It also has a certain reference significance for the future exploration of the performance of neural network.

关 键 词：PERLIN噪声 黑盒攻击 神经网络 鲁棒性 

分 类 号：TN919.81[电子电信—通信与信息系统]