Tensor-Based GAN to Defense Adversarial Attacks for Cyber-Physical-Social System

In recent years, considerable achievements made by scientists in artificial intelligence have provided strong support for the development of Cyber-Physical-Social-Systems. Meanwhile, deep computation to realize AI is widely used in various applications, such as object classification. However, in classification application, the nice perturbations carefully designed added to the input data will make the classification label wrong, and the difference between the adversarial sample and the original one is hard to detect for humans. Hence, this paper proposed a defense mechanism which was tensor-based deep convolution generation adversarial network model to resist the possible attacks of classifiers in CPSSs. The method could play a good defense effect of the application, because of the reconstruction method using a complete Tensor-DCGAN model. The decoder with the same structure as the discriminator was used to decode the classified data, and the generator was used to generate a data which was the most similar to the data to be classified. The experimental results shown that our improved method not only preserved the advantages of the original model, but also reduced the parameters required by the original methods, and be more importantly, it made up the shortcomings of the original of color images and network data.