Hanyong Lee

Ph.D. student in artificial intelligence at CAU AutoML Lab.

7 minute read

Summary: This study proposes the BitAbuse dataset, which comprises real phishing cases containing visually perturbed (VP) texts. The dataset aims to enhance the performance of language models and support adversarial attack defense research.

1 Introduction

image

  • Social engineering attacks exploit victims’ psychological vulnerabilities to extract confidential information.
    • Attack types: phishing, spam, pretexting, baiting, tailgating, etc.
  • Phishing attacks target victims through text-based communication, such as emails and SMS.
    • Visually perturbed (VP) texts are used to bypass security systems.
    • Example of VP text: "Bitcoin" ➔ "ßitcöın"
  • VP text-based phishing attacks can be mitigated by restoring the original text.
    • Existing research primarily focuses on restoration techniques.
    • However, limitations exist in models like Viper, which alter unrealistic fixed components.
  • LEGIT research generated a VP text dataset while considering readability.
    • However, research on real-world VP texts remains insufficient.
  • Defense systems based on synthetic datasets may pose risks due to real-world discrepancies.
    • Solution: Combining real VP texts with synthetic texts.
  • Proposed dataset: BitAbuse
    • Constructed from 262,258 phishing-related emails.
    • Contains 26,591 VP sentences and 298,989 non-VP English sentences.
    • Introduces three datasets: BitCore, BitViper, and BitAbuse.
    • A pilot study was conducted to analyze dataset characteristics.
  • BitAbuse phishing attack dataset is publicly available.

2 Related Work

  • Challenges in VP text research:
    • Difficulties in obtaining sufficient data due to its presence in spam emails.
    • Lack of datasets reflecting real-world phishing attack scenarios.
    • Existing datasets are only valid under specific conditions (e.g., internationalized domain names).
    • Traditional research integrates datasets for VP text restoration techniques.
  • Two notable VP text dataset integration studies:
    • TextBugger:
      • Generates VP texts using predefined homograph pairs and transformation methods.
      • Selects characters in text and replaces them with VP characters to degrade LMs’ performance.
      • Useful for exposing vulnerabilities in security-sensitive tasks like sentiment analysis and malicious content detection.
    • Viper:
      • Searches for homographs and generates VP texts based on embedding techniques.
      • Replaces text characters with VP characters and induces visual interference based on replacement probabilities.
  • Traditional VP text restoration methods:
    • Utilize SimChar DB, OCR, spell checker, and LMs.
    • SimChar DB automatically collects homographs from Unicode character sets, detects VP characters, and restores text using predefined restoration tables.
    • OCR-based methods detect phishing attacks by inserting VP characters into IDNs.
    • Spell checkers restore malicious texts distributed across social networks.
    • Combining two LMs (BERT & GPT) for restoration strategies is also considered.
  • Common limitations in previous research:
    • Existing phishing attack restoration datasets lack real VP texts.
    • This leads to overestimation or underestimation of restoration performance in real-world scenarios, potentially resulting in unstable pre-trained LM models.
  • This study collects VP texts used on bitcoinabuse[.]com to construct a new dataset for phishing attack research.

3 BitAbuse

image

  • Email data collection from bitcoinabuse[.]com to obtain VP texts used in phishing attacks.
  • Bitcoin Abuse website serves as a platform where users worldwide report Bitcoin-related scams.
  • The platform allows users to upload anonymized phishing emails, enabling safe data collection.
  • A total of 262,258 phishing-related emails were collected from May 16, 2017, to January 15, 2022.
  • The goal was to construct an English VP text dataset, but non-English texts were also included, requiring irrelevant email filtering.
  • A BERT model was trained to classify English texts, using 16,598 manually labeled samples.
  • Final dataset: 178,054 English emails retained for further processing.
  • Sentences were split into a maximum length of 512 tokens, resulting in 326,732 sentences.
  • Regular expressions were used to remove unnecessary character sequences.
  • Manual annotation of VP texts was conducted on 326,732 sentences.
  • To reduce inefficiencies, VP words were converted to non-VP word labels for automation.
  • 1,152 irrelevant sentences were removed.
  • Final datasets: BitCore, BitViper, and BitAbuse.

4 Experimental Settings

  • Experimental databases & methodologies
    • Methods used: SimChar DB (Suzuki et al., 2019), OCR (Sawabe et al., 2019), Spell Checker (Imam et al., 2022), Character BERT (El Boukkouri et al., 2020), GPT-4o mini (OpenAI, 2023)
    • Evaluation metrics: Word Level Accuracy, Word Level Jaccard, BLEU
  • Restoration performance evaluation
    • Five different restoration methods were evaluated:
      1. SimChar DB-based: Identifies alphabet homographs and restores them.
      2. OCR-based: Applies OCR to each character and selects the most probable character.
      3. Spell Checker-based: Splits sentences into words and uses Levenshtein Distance for restoration.
      4. Character BERT: Processes token sequences at the character level for context-based restoration.
      5. GPT-4o mini: Evaluates performance using the latest large language model (LLM).
  • Character BERT settings
    • Learning rate: 5×10⁻⁵, Batch size: 32, Epochs: 10
    • Optimizer: AdamW (β₁ = 0.9, β₂ = 0.999, weight_decay = 0)
    • Input & output: Character-level token sequences.
  • GPT-4o mini model access
    • Conducted via OpenAI’s inference API with a carefully designed prompt.
  • Evaluation metrics
    • Word Level Accuracy: Measures if the restored word matches the target word at each position.
    • Word Level Jaccard: Computes the ratio of the intersection and union of word sets in predicted vs. labeled sentences.
    • BLEU score: Calculates n-gram precision between predicted and labeled sentences.
  • Data split
    • BitAbuse dataset split into 60% training, 20% validation, and 20% testing.
    • Each method’s performance was evaluated on the test set, averaging results from 10 random train-test splits.

5 Experimental Results

image

  • Exploratory Data Analysis: Analyzed VP words, VP characters, and their ratios to aid the development of phishing attack defense methodologies.

image

  • VP Sentence Histogram: Provided histograms illustrating the occurrence rates of VP characters based on sentence length across BitCore, BitViper, and BitAbuse datasets.

image

  • VP Character-Word Association Graph: Visualized the clustering of VP characters and words using the Yifan Hu algorithm, showing vowels as central elements in key clusters.

image

  • Restoration Performance Comparison: Evaluated the restoration performance of SimChar DB, OCR, Spell Checker, Character BERT, and GPT-4o mini-based methods, with Character BERT outperforming the others.

image

  • VP Word Restoration Errors: Character BERT-based methods exhibited increased failures when restoring continuous VP characters.

image

  • Word-Level Accuracy Evaluation: Performance of Character BERT-based methods was evaluated across three datasets based on VP character ratios, with BitCore achieving the strongest performance.

image

  • Performance Variations by Training Volume: Observed performance degradation in BitViper and BitAbuse datasets when trained with only 1% or 5% of VP sentences.

  • Model Generalization Capability: Despite a low training data ratio, the model successfully restored text, demonstrating practical advantages through rapid training.

  • Additional Experimental Results Summary: Concluded that Character BERT-based methods demonstrated relatively superior performance, and a sufficient number of VP sentences is essential for building stable models.

6 Discussion

  • Comparison of restoration methods based on VP character ratio:
    • Character BERT-based methods effectively improved performance as VP character ratios increased.
    • Spell Checker-based methods suffered a steep decline in performance due to their inability to utilize contextual information.
    • GPT-4o mini-based methods exhibited degraded performance as they failed to maintain input-output sequence order and indexing.
  • Jaccard and BLEU Performance Evaluations:
    • Jaccard scores closely correlated with Word Level Accuracy.
    • BLEU scores were more sensitive to contextual accuracy, allowing simpler methods that preserve structure and meaning to achieve higher scores.
  • Comparison of five restoration methods across three datasets:
    • Character BERT-based methods clearly outperformed all other approaches.
    • SimChar DB-based methods were limited to restoring a single VP character to only one non-VP character.
  • VP Character Restoration Capabilities of Each Method:
    • Character BERT-based methods learned context directly and performed fast restoration.
    • GPT-4o mini exhibited poor performance when multiple VP characters were present.
    • Safety features caused error responses when handling unethical content.
  • Character BERT-based method achieving nearly 100% accuracy on BitCore dataset:
    • Suggests potential applications in digital forensics and secure messaging systems.
    • The model could help address critical challenges in secure communication.

7 Conclusion

  • Developed three VP text datasets: BitCore, BitViper, and BitAbuse.
  • BitCore and BitViper exhibit significantly different characteristics, with LMs-based restoration methods demonstrating strong robustness and potential across all datasets.
  • BitAbuse is available for download, pre-trained with 325,580 VP sentences.
  • Future research should explore hybrid approaches combining OCR and Character BERT.
  • Internalizing data within LMs could mitigate the excessive data consumption problem.
  • Keyword-based tendencies may facilitate the development of lightweight yet accurate LMs for real-world phishing attacks.
  • Zero-shot performance validation of the BitAbuse model is also necessary.

8 Limitations

  • The VP text restoration experiments in this study did not include additional restoration methods to avoid exceeding the research scope.
  • No direct comparison was made between Character BERT-based methods and other LM-based restoration techniques, making it difficult to assess Character BERT’s superiority.
  • BitAbuse dataset contains only Bitcoin scam data, limiting its ability to reflect diverse phishing attack scenarios.
  • Phishing attacks are likely to evolve over time, increasing in variety and complexity; failing to account for this diversity may reduce generalization capability.
  • There is a risk that non-experts could misuse the dataset to learn and execute phishing attacks.
  • Misuse of the BitAbuse dataset may lead to more sophisticated phishing attacks, increasing victim exposure.
  • Although the dataset and model are publicly available, there is no clear regulation against their technological misuse.

9 Ethics Statement

  • The dataset created in this study is intended for phishing attack defense research.
  • However, non-experts may use the dataset to learn and execute phishing attacks.
  • It could be exploited for developing malicious tools like WormGPT from the dark web or PoisonGPT released by Mithril Security.
  • This may lead to more sophisticated phishing attacks and increased victims.
  • Legal responsibility cannot be imposed for damages caused by misuse of the dataset.
  • Many countries lack clear regulations against dataset misuse, requiring careful consideration and monitoring.
  • Although the dataset and model used in this study are publicly available, they should not be used for non-research purposes.

Reader Comments

  • This paper was written by the author.
  • The primary focus is on introducing the dataset, and methodologies for solving the task are limited to basic applications of language models.
  • Future research could incorporate multimodal approaches by utilizing character shape information to enhance performance.
  • However, even without multimodal approaches, LLMs can achieve sufficiently high performance, assuming cost is not a constraint.

Comments

You May Also Enjoy