A Neural Attack Model for Cracking Passwords in Adversarial Environments

In many scenarios, one has to enter her text or graphical password in a public area, such as unlocking the smartphone on the street, and entering the password when she pays with a debit card in a shopping mall. However, the environment where the password is entered may be adversarial as it is almost impossible to prevent adversaries from premeditated installation of surveillance and/or eavesdropping equipment in public areas. In this work, we investigate password security in such extreme adversarial environments in which every single interaction between humans (provers) and input terminals (verifiers) is transparent to the attacker. We first present a neural network-based attack model, which consists of a feature extraction model and a prediction model. Experimental results show that the neural model attains an accuracy of more than 80% in password prediction in three real-world authentication systems. We also propose a risk alert system based on the attack model. It can issue a timely warning notice when the password in use is at high security risk.