基于类别重构损失检测音频对抗样本的胶囊网络

I systems are also suffering from security threats while serving human beings. Among them, anti-sample attacks play a major role. Researchers have carried out various ways of research in order to eliminate the impact of anti-sample attacks. The capsule network can be verified by setting a global threshold θ， It is used to compare the reconstruction error between the input sample and the reconstruction sample to detect the counter sample, but the global threshold θ There is a certain problem, that is, it is not accurate to determine different types of confrontation samples. Therefore, a class-based reconstruction loss mean detector is proposed, and an audio confrontation sample attack algorithm based on the prediction ability of echo state network is proposed. Finally, the experimental verification of the proposed new detector\'s detection ability of confrontation samples, and also verified the detection ability of capsule network for audio confrontation samples.