Classifiers Protected against Attacks by Fusion of Multi-Branch Perturbed GAN

Deep learning is widely used in classification tasks to achieve advanced performance. However, in the face of well-designed image classifications, such as the Fast Gradient Sign Method (FGSM), there are glaring errors. This paper proposes a new technique to eliminate interference using generative adversarial networks (GAN), called multi-branch perturbed generative adversarial networks ( MBP-GAN). MBP-GAN minimizes the input feature flow graph in generator noise filtering by introducing multi-branch fusion perturbations. This makes the sample of the generator more aware of this perturbation, thereby improving the ability of the generator and discriminator to resist classification against attacks in combat training. Through this kind of training, this model can be used as a defense against arbitrary attacks. Then we design the loss function, so that the generator and the discriminator can keep accurate results for general images and classification against images. We verify our experimental results on the MNIST, F-MNIST and CelebA datasets. The results show that the MBP-GAN can effectively eliminate the interference from the classification against the attack.