DenseGAN: A Password Guessing Model Based on DenseNet and PassGAN

Abstract

Password authentication has become one of the most significant authentication methods because of low cost and convenience, and its security is getting more and more attention. The vulnerability of password security mainly lies in the password construction method which inevitably has many human characteristics. With the development of deep learning, these human characteristics are more and more explored, which bring new challenges to password security. In 2019, a PassGAN password guessing model was proposed, and its performance is remarkable when the maximum training password length is 10. However, when the length is extended to 15, the performance gets worse. To address this issue, in this paper an approach is proposed to innovate the structure of PassGAN by using DenseNet, and two novel password guessing DenseGAN models are proposed, which both can generate high-quality password guesses. With the first DenseGAN model, when the maximum training password length is 15, the generated passwords were able to match 2.7–4.8% of the passwords in the testing datasets more than PassGAN. Specifically, with the second DenseGAN model, when the maximum training password length is 10, the generated passwords were able to match 0.5% of the passwords in the testing datasets more than PassGAN, when the maximum training password length is 15, the match is 6.2% to 12.5% of the passwords more than PassGAN.