GNPassGAN: Improved Generative Adversarial Networks For Trawling Offline Password Guessing

Fangyi Yu, Miguel Vargas Martin

TL;DR

This paper introduces GNPassGAN, an improved Generative Adversarial Network (GAN) for offline password guessing. By incorporating gradient normalization and modifying the GAN architecture, GNPassGAN significantly outperforms the previous state-of-the-art PassGAN, achieving higher password guessing accuracy and generating fewer duplicate passwords.

What Problem Does This Paper Address?

The security of user accounts heavily relies on passwords, yet many users create weak passwords due to common, predictable patterns. Simulating realistic password guessing attacks is crucial for understanding password security and developing better defenses. Traditional methods like dictionary attacks require domain expertise to configure effectively. While deep learning offers a way to learn password patterns without explicit rules, existing models like PassGAN still face challenges. This paper aims to improve the effectiveness and efficiency of deep learning-based password guessing, specifically for trawling offline attacks, which involve guessing passwords from a large pool of potential users without prior knowledge of their identities.

Methods

The core of this work is GNPassGAN, an enhancement of the PassGAN model. The authors introduce three key modifications:

  • Gradient Normalization: Instead of Wasserstein GAN's (WGAN) weight clipping or gradient penalty, GNPassGAN employs gradient normalization in the discriminator. This technique, proven effective in computer vision GANs, aims to stabilize training and improve the discriminator's capacity.
  • Generator Activation Function: The generator's final activation function is changed from softmax to tanh.
  • Loss Function: The authors switch from Wasserstein loss to a binary entropy loss within a sigmoid layer.

These architectural and loss function changes are designed to address issues like mode collapse and improve the overall training stability and performance of the GAN.

Datasets

The experiments utilize publicly available password datasets:

  • Training: The rockyou dataset was used for training the models.
  • Testing: The phpbb dataset and a disjoint subset of the rockyou dataset were used for testing.

The paper also analyzes password distributions based on length, noting that most real-world passwords are less than or equal to 10 characters, but also considers a "Char812" scenario (passwords between 8 and 12 characters) to better reflect modern password policies.

Key Insights and Ideas

  • Gradient Normalization's Impact: The adoption of gradient normalization, a technique previously successful in computer vision, proves beneficial for password guessing GANs, leading to improved performance and stability.
  • Cross-Site Generalization: GNPassGAN demonstrates good performance even when tested on a dataset (phpbb) different from the training set (rockyou), suggesting its ability to generalize to cross-site password guessing scenarios.
  • Password Length Matters: The study confirms that longer passwords are significantly harder to guess, highlighting the importance of minimum password length policies.
  • Honeyword Potential: Unmatched generated passwords from GNPassGAN are plausible candidates for "honeywords" – decoy passwords designed to detect data breaches. This opens up a new avenue for security research.

Key Results

GNPassGAN significantly outperforms PassGAN:

  • Guessing Accuracy: When generating 108 passwords for the "Char10" scenario (passwords ≤ 10 characters), GNPassGAN matches 88.03% more passwords than PassGAN. For the "Char812" scenario (passwords 8-12 characters), it matches six times more passwords.
  • Duplicate Generation: GNPassGAN generates 31.69% fewer duplicates in the "Char10" scenario and 61.80% fewer duplicates in the "Char812" scenario compared to PassGAN. This indicates a better resolution of the mode collapse problem.
  • Training Iterations: The study suggests that optimal performance for both models is achieved around the 180,000th to 200,000th training iteration, with GNPassGAN showing strong performance even before reaching the full 200,000 iterations.

Limitations and Future Work

  • Password Policy Realism: The current experiments, like much prior work, primarily focus on passwords less than or equal to 10 characters. Future work should more rigorously incorporate real-world password policies (minimum length, character types) into dataset preprocessing and model evaluation.
  • Hybrid Models: The authors suggest combining GNPassGAN with rule-based models (like HashCat) to create more powerful hybrid password guessing tools.
  • Honeyword Generation: Further research is planned to explore GNPassGAN's capability in generating high-quality honeywords and quantitatively evaluate their effectiveness in detecting password breaches.

Summary updated

Abstract

The security of passwords depends on a thorough understanding of the strategies used by attackers. Unfortunately, real-world adversaries use pragmatic guessing tactics like dictionary attacks, which are difficult to simulate in password security research. Dictionary attacks must be carefully configured and modified to represent an actual threat. This approach, however, needs domain-specific knowledge and expertise that are difficult to duplicate. This paper reviews various deep learning-based password guessing approaches that do not require domain knowledge or assumptions about users' password structures and combinations. It also introduces GNPassGAN, a password guessing tool built on generative adversarial networks for trawling offline attacks. In comparison to the state-of-the-art PassGAN model, GNPassGAN is capable of guessing 88.03% more passwords and generating 31.69% fewer duplicates.
BibTeX
@inproceedings{yu2022gnpassgan,
  author = {Fangyi Yu and Miguel Vargas Martin},
  title = {{GNPassGAN: Improved Generative Adversarial Networks For Trawling Offline Password Guessing}},
  booktitle = {EuroS\&P Workshops},
  year = {2022},
  url = {https://openreview.net/forum?id=Rj0gwh0mqg},
}
Figure