PG-RNN:一种基于递归神经网络的密码猜测模型

PDF下载HTML

Title:

PG-RNN: a password-guessing model based on recurrent neural networks

作者:

滕南君^1,2, 鲁华祥^1,3,4, 金敏¹, 叶俊彬^1,2, 李志远^1,2

1. 中国科学院 半导体研究所, 北京 100083;
2. 中国科学院大学, 北京 100089;
3. 中国科学院 脑科学与智能技术卓越创新中心, 上海 200031;
4. 半导体神经网络智能感知与计算技术北京市重点实验室, 北京 100083

Author(s):

TENG Nanjun^1,2, LU Huaxiang^1,3,4, JIN Min¹, YE Junbin^1,2, LI Zhiyuan^1,2

1. Institute of Semiconductors, Chinese Academy of Sciences, Beijing 100083, China;
2. University of Chinese Academy of Sciences, Beijing 100089, China;
3. Center for Excellence in Brain Science and Intelligence Technology, Chinese Academy of Sciences, Shanghai 200031, China;
4. Semiconductor Neural Network Intelligent Perception and Computing Technology Beijing Key Lab, Beijing 100083, China

关键词:

密码生成; 深度学习; 递归神经网络; Markov; 密码猜测

Keywords:

password generation; deep learning; recurrent neural networks; Markov; password guessing

分类号:

TP391

DOI:

10.11992/tis.201712006

摘要:

用户名—密码（口令）是目前最流行的用户身份认证方式，鉴于获取真实的大规模密码明文非常困难，利用密码猜测技术来生成大规模密码集，可以评估密码猜测算法效率、检测现有用户密码保护机制的缺陷等，是研究密码安全性的主要方法。本文提出了一种基于递归神经网络的密码猜测概率模型（password guessing RNN， PG-RNN），区别于传统的基于人为设计规则的密码生成方法，递归神经网络能够自动地学习到密码集本身的分布特征和字符规律。因此，在泄露的真实用户密码集上训练后的递归神经网络，能够生成非常接近训练集真实数据的密码，避免了人为设定规则来破译密码的局限性。实验结果表明，PG-RNN生成的密码在结构字符类型、密码长度分布上比Markov模型更好地接近原始训练数据的分布特征，同时在真实密码匹配度上，本文提出的PG-RNN模型比目前较好的基于生成对抗网络的PassGAN模型提高了1.2%。

Abstract:

Passwords are the most popular way of user ID authentication. However, it is rather difficult to obtain large-scale real text passwords. Generating large-scale password sets based on password-guessing techniques is a principal method to research password security, which can be applied to evaluate the efficiency of password-guessing algorithm and detect the defects of existing user-password protective mechanisms. In this paper, we propose a password guessing-based recurrent neural network (PG-RNN) model. Our model can directly and automatically infer the distribution characteristics and character rules from the data of password sets, which is different from the traditional password generating method based on manual design rule. Therefore, an RNN model that has been trained on a disclosed real user password set can generate passwords very close to the real data of the training set, which avoids the limitations of manual setting for password guessing. The results of our experiments show that PG-RNN can generate passwords closer to primitive data distribution more than Markov in password length and character structure categories. When evaluating on large password dataset, the proposed PG-RNN model matching outperforms that of PassGAN, which is based on generative adversarial networks, by more than 1.2%.