Abstract
The password has become today's dominant method of authentication. While brute-force attack methods such as HashCat and John the Ripper have proven impractical, research has shifted toward password guessing. State-of-the-art approaches such as the Markov Model and probabilistic context-free grammar (PCFG) are based on statistical probability and require significant computation. Neural networks have proven more accurate and practical, yet raw models often fail in cross-site attacks because each dataset contains unique, site-specific features. In this paper, we propose GENPass, a multi-source deep learning model for generating "general" passwords. GENPass learns from several datasets and ensures the output wordlist maintains high accuracy across different platforms using adversarial generation. Our core generator, PCFG+LSTM (PL), is the first to combine neural networks with PCFG. Compared with standard Long Short-Term Memory (LSTM) models, PL increases matching rates by 16%–30% in cross-site tests. Furthermore, we introduce GENPass-pro, which utilizes probabilistic weighting to further enhance performance.
1. Introduction
In cybersecurity, the shift from brute-force attacks to intelligent password guessing reflects the reality of human behavior: passwords are rarely random. However, existing deep learning models suffer from a "site-specificity" problem. A model trained on a leak from one site (e.g., RockYou) often overfits to that specific user base, performing poorly when attempting to guess passwords for a different site (e.g., LinkedIn).
GENPass addresses this challenge by creating a framework that learns the underlying, universal logic of password creation across multiple platforms. By generalizing features from diverse sources, GENPass produces wordlists that are more effective for cross-site attacks than models trained on single or simply aggregated datasets.
2. Methodology
2.1 The PCFG+LSTM (PL) Model
Standard LSTMs predict passwords character-by-character, which often results in structurally invalid "gibberish." GENPass introduces the PL model, which uses PCFG to pre-process passwords into structural tags. For example, the password password123 is converted into the tag L8 D3 (8 letters followed by 3 digits). The LSTM then learns these tagged sequences, effectively moving from character-level to "word-level" training. This ensures the model preserves the semantic structures humans naturally favor.
2.2 Adversarial Generality and Filtering
To prevent the model from memorizing a single dataset, GENPass employs multiple PL generators, each trained on a different source. These generators feed into a CNN Classifier and a Discriminator. The classifier attempts to identify which source dataset a generated password originated from. The discriminator only accepts passwords that the classifier cannot easily attribute to a specific site. This adversarial process forces the model to produce "general" passwords that share features across all training sets.
2.3 GENPass-pro: Probabilistic Weighting
An upgraded version, GENPass-pro, incorporates Kullback-Leibler (KL) divergence to assign weights to different source datasets. This recognizes that some leaks are more representative of "typical" human behavior than others, allowing the model to prioritize higher-quality data sources during the generation phase.
3. Experimental Setup
The study utilizes four major real-world leaked password datasets to evaluate performance:
- MySpace: 37,144 passwords.
- phpBB: 184,389 passwords.
- RockYou: ~14.3 million passwords.
- LinkedIn: ~60.1 million passwords.
The evaluation focused on English-language environments, as different linguistic habits (such as Chinese Pinyin) require distinct transfer learning techniques.
4. Results and Discussion
4.1 Efficiency and Matching Rates
The PL model demonstrates massive efficiency gains over traditional LSTMs. To achieve a 50% matching rate on the MySpace dataset, a standard LSTM required $10^{12}$ guesses, while the PL model reached the same threshold with only $10^7$ guesses—a five-order-of-magnitude improvement. In cross-site tests, GENPass achieved matching rates 20% higher than models trained by simply mixing datasets.
4.2 High-Volume Performance
When generating large-scale wordlists (exceeding $10^{10}$ passwords), GENPass-pro showed a further 2%–3% improvement over the standard GENPass model. It reached a 36.2% matching rate on RockYou and 23.8% on LinkedIn, proving its robustness in high-volume attack scenarios.
4.3 The Failure of Simple Mixing
A key insight from the research is that simply merging datasets into a single training file actually decreases performance. Large, noisy datasets tend to drown out the useful patterns found in smaller, more specific leaks. GENPass’s multi-generator approach successfully avoids this "data dilution."
5. Conclusion
GENPass represents a significant advancement in password guessing by combining structural grammar with adversarial deep learning. By focusing on generality rather than site-specific memorization, it provides a more practical tool for cross-site security auditing. Future work will explore applying these techniques to other domains, such as identifying vulnerabilities in smart grids and intelligent transportation systems, while also expanding the model's linguistic capabilities through transfer learning.
Summary updated