Structural Rule-Driven Password Recovery with Multi-GPU Parallelization Techniques

This paper proposes a novel password guessing method that produces structural sets from password training sets, which fundamentally differs from conventional approaches that generate dictionaries. This paper further proposes a parallel method to generate candidate passwords from a structured set on multiple GPUs. In the training phase, each password in the training set is segmented based on the pre-built corpus. The segmentation with maximum coverage and the minimum number of segments is adopted as the canonical structure representing the corresponding password. Then the structural set for the training password set is selected based on the number of guesses required. During password recovery, complex structures are split into two parts. All combinations of each parts are generated on the CPU and then they are concatenated to build guesses on the GPU. This design maximizes data parallelism and thread-level parallelism in multicore CPUs while implementing software pipelining for efficient CPU-GPU collaboration. Experimental results on a $4 \times$ RTX 3080 GPU server demonstrate remarkable improvements: achieving $457 \times$ faster hashing speed for the MD5 hash algorithm and increasing crack rates from 21 % to $\mathbf{5 1 \%}$ compared to the PassGPT-generated dictionary. For other high-speed hashing algorithms whose computational throughput exceeds the I/O bandwidth of the disk subsystem, our method can generate more candidate passwords at the same time to significantly improve cracking success rates.