A Large-scale Analysis of the Mnemonic Password Advice

How to choose a strong but still easily memorable password? An often recommended advice is to memorize a random sentence (the mnemonic) and to concatenate the words’ initials: a so-called mnemonic password. The paper in hand analyzes the effectiveness of this advice—in terms of the obtained password strength—and sheds light on various related aspects. While it is infeasible to obtain a sufficiently large sample of human-chosen mnemonics, the password strength depends only on the distribution of certain character probabilities. We provide several pieces of evidence that these character probabilities are approximately the same for human-chosen mnemonics and sentences from a web crawl and exploit this connection for our analyses. The presented analyses are independent of cracking software, avoid privacy concerns, and allow full control over the details of how passwords are generated from sentences. In particular, the paper introduces the following original research contributions: (1) construction of one of the largest corpora of human-chosen mnemonics, (2) construction of two web sentence corpora from the 27.3 TB ClueWeb12 web crawl, (3) demonstration of the suitability of web sentences as substitutes for mnemonics in password strength analyses, (4) improved estimation of password probabilities by position-dependent language models, and (5) analysis of the obtained password strength using web sentence samples of different sentence complexity and using 18 generation rules for mnemonic password construction. Our findings include both expected and less expected results, among others: mnemonic passwords from lowercase letters only provide comparable strength to mnemonic passwords that exploit the 7-bit visible ASCII character set, less complex mnemonics reduce password strength in offline scenarios by less than expected, and longer mnemonic passwords provide more security in an offline but not necessarily in an online scenario. When compared to passwords generated by uniform sampling from a dictionary, distributions of mnemonic passwords can reach the same strength against offline attacks with less characters.