Botnet detection and information leakage mitigation with differential privacy under generative adversarial networks

Botnets are a serious threat to computer networks. New botnets are created with the aim of evading detection by making modifications. The proposed methods are insufficient for detecting modified botnets. A Generative Adversarial Network (GAN) trained with generated and real data was used to improve detection performance against altered botnets. GANs remember training samples, which leads to the leakage of critical information. Differential Privacy (DP) was utilized to mitigate information leakage by adding noise to the gradients during the training process. Due to instability in the training of the GAN and the effect of the DP method on reducing stability, the mixup method was proposed to stabilize the GAN training, which generates new samples through the linear interpolation of multiple samples. The efficiency of the proposed method in modified botnet detection and information leakage mitigation was acceptable compared to other methods and achieved a high classification accuracy of 97.4%.