Authors - Aaqib Hakeem, Akshay V, Parthav Mathu, Kotnada Yogesh, Gokul Kannan Sadasivam Abstract - Passwords remain one of the most widely deployed authentication mechanisms despite well-documented vulnerabilities to guessing attacks. Recent deep learning approaches, including Password Guessing using Temporal Convolutional Networks (PGTCN), have demonstrated that sequence modeling can effectively capture structural regularities in leaked password corpora. However, practical performance often depends not only on model architecture but also on training stability, batching strategy, and decoding configuration. In this work, we investigate a partition-aware training and generation pipeline built around a single Temporal Convolutional Network (TCN). Rather than introducing additional architectural complexity, the proposed framework emphasizes standardized preprocessing, balanced data partitioning for stable batching, optimized training procedures, and large-batch probabilistic decoding. A lightweight buffering layer is incorporated to decouple generation from evaluation and improve throughput without requiring distributed training infrastructure. Experiments on multiple real-world leaked password datasets show consistent, though modest, improvements in match rate compared to the PGTCN baseline under same-site evaluation. The results suggest that careful optimization and pipeline-level design can yield measurable gains in candidate ordering while maintaining reproducibility and implementation simplicity.
Open Zoom