Neural networks trained to learn attackers’ approaches to brute-force password guessing can be used as a way to enforce minimal password security without resorting to large blocklists and cumbersome combinations of letters, numbers, and special symbols, a research team at Carnegie Mellon University conclude in a new paper.
Using a neural network model built into a password-strength meter and recruiting users through Amazon’s Mechanical Turk, the researchers at CMU’s CyLab Security and Privacy Institute evaluated a series of different password recommendations, from eight-character passwords using a single class (letters, for example) to 16-character passwords using four classes — lowercase letter, uppercase letter, numbers, and symbols — as well as different blocklists. The researchers found that just requiring 12 characters of a single class and meeting the neural network’s recommendations resulted in hard-to-crack passwords that should be sufficient for most uses.
Interestingly, requiring that users combine different cases, numbers, and symbols is not necessary, says Lujo Bauer, professor of electrical and computer engineering at CMU’s Institute for Software Research. Attackers’ current tools have become pretty good at guessing passwords that consist of the four classes of characters, making any benefit marginal, he says.
“In part, because previously there were many fewer three- and four-class passwords that had been leaked and were available to attackers, [it had been] harder for attackers to develop ways of guessing those passwords effectively,” Bauer says. “Now that there have been many such passwords leaked, it’s much easier to train an algorithm to guess them.”
The research is about finding the best balance between usability and security for passwords. The researchers’ neural network models attackers to determine what can be easily guessed by current methods. Although “easily guessed” is relative — the final neural network models whether an attacker could find the password if that person had a list of 10 billion possibilities.
The conclusion is that websites and other services can simplify their requirements for users, the researchers say in their paper.
“Our experimental results provide the first concrete evidence that character-class requirements should be avoided not only because users tend to find them annoying, but also because they don’t provide substantial benefit against attackers using state-of-the-art password-cracking tools,” the researchers state. “[A]n expert attacker can guess (single-class, two-class, or four-class) passwords with equal success rates.”
The team includes Joshua Tan, Lujo Bauer, Nicolas Christin, and Lorrie Faith Cranor, all of Carnegie Mellon University.
The work is based on previous Carnegie Mellon research that created a password-strength meter using a neural network trained with a certain size of password dictionary. While typical blocklists might include a large number of dictionary words and combinations that include numbers and special characters, the small size of the neural network model — a few hundred kilobytes of memory — means it can be directly integrated into a webpage.
The combination of a minimum password length and the requirement that the password meet the neural networks’ strength standard can help companies ensure that their employees are creating hard-to-break passwords, CMU’s Bauer says.
“The choice of minimum-strength threshold depends on security requirements,” the researchers state in the paper. “Too low a threshold may not provide enough defense — particularly against online attacks — and too high a threshold may unacceptably inhibit usability.”
In addition, companies should require that workers use two-factor authentication as a way to improve security and foil offline attackers, Bauer says.
“It’s not as painful as people think, as we reported on in another paper, and enforces a password policy that prevents using leaked — or at least common — passwords,” he says. “Emphasize the importance of not reusing passwords and encourage using password managers.”
Veteran technology journalist of more than 20 years. Former research engineer. Written for more than two dozen publications, including CNET News.com, Dark Reading, MIT’s Technology Review, Popular Science, and Wired News. Five awards for journalism, including Best Deadline … View Full Bio