基于双边滤波与自编码器的对抗样本防御方法

Anti-Sample Defense Method Based on Bilateral Filtering and Autoencoder

深度学习在图像分类等领域被广泛应用的同时,也存在着对抗样本攻击的问题。针对这一问题,提出基于双边滤波与卷积降噪自编码器的对抗样本防御方法BF-CDAE。首先对加入对抗扰动的图片进行双边滤波,初步去除图片的噪音扰动,然后将去噪后的图片送入卷积降噪自编码器中,对高维数据进行特征提取,进一步去除对抗样本中误导模型识别的噪音。实验结果表明,在不影响原图像分类结果的同时,针对采用FGSM方法生成的对抗样本攻击,利用该防御方法可将分类准确率恢复到93.14%,证明了该防御方法的有效性。

While deep learning is widely used in image classification and other fields,it also has the problem of resisting sample attacks.In response to this problem,an adversarial sample defense method BF-CDAE based on bilateral filtering and convolutional noise reduction autoencoder is proposed.Firstly,bilateral filtering is performed on the pictures that are added to the anti-disturbance,and the noise disturbance of the picture is initially removed,and then the denoised pictures are sent to the convolutional denoising autoencoder,and the high-dimensional data is feature extracted to further remove the misleading in the counter-sample model recognition noise.The experimental results show that,while not affecting the original image classification results,the use of this defense method can restore the classification accuracy to 93.14%in the attack based on the adversarial sample generated by the FGSM method,which proves the effectiveness of the defense scheme.