Prob-Hashcat: Accelerating Probabilistic Password Guessing with Hashcat by Hundreds of Times

While the academic community has proposed dozens of probabilistic password guessing models to improve the success rate of password guessing, few studies have considered the speed of generating password guesses (which is a crucial factor in realistic password guessing scenarios). Real-world attackers often aim to crack more passwords in less time, and the speed of these models thus becomes a significant concern. Consequently, real-world attackers tend to prefer simple heuristic methods (such as Rule attack and Mask attack) and off-the-shelf password cracking tools (such as Hashcat and John the Ripper), over academic probabilistic password guessing models, despite the latter’s superior scientific flavor. To fill this gap, we introduce an offline guessing speed theory for measuring the speed of generating guesses, and a probabilistic model parallelization framework to accelerate probabilistic guessing models. We use our theory to elucidate the acceleration principles of our framework, and provide a method for integrating this framework with Hashcat to fully leverage GPU acceleration. Our framework is scalable, generalizable for various probabilistic models and applicable to guessing scenarios that tackle multiple hashes. To exhibit the practical value of our theory and framework, we implement two probabilistic password guessing models, PCFG and OMEN, within our framework and develop a high-speed cracking tool, Prob-hashcat. Extensive experiments on eight large real-world password datasets demonstrate the effectiveness of Prob-hashcat: (1) Both models can generate over 100 million guesses per second on a common computer and handle computing and matching hashes; (2) The average speeds of PCFG and OMEN increase by 31104 times and 213646 times, respectively, compared to their original speeds; (3) Within 20 minutes, the cracking rates of PCFG and OMEN increase by 12%42% and 16%182%, respectively, compared to their original rates, and this advantage becomes larger as the cracking time increases. We, for the first time, addresses the issue of employing GPU acceleration for probabilistic password models with Hashcat, highlighting that these models are significantly greater threats in real-world guessing attacks than expected.