Towards Performance of NLP Transformers on URL-Based Phishing Detection for Mobile Devices
Why this work is in the frame
A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.
Bibliographic record
Abstract
ackers are increasingly launching phishing attacks via SMS and social media. Games and dating apps introduce yet another attack vector. However, current deep learning-based phishing detection applications do not apply to mobile devices due to the computational burden. We propose a lightweight phishing detection algorithm that distinguishes phishing from legitimate websites solely from URLs to be used in mobile devices. As a baseline performance, we apply Artificial Neural Networks (ANNs) to URL-based and HTML-based website features. A model search results in 15 ANN models with accuracies >96%, comparable to state-of-the-art approaches. Next, we test the performance of deep ANNs on URLbased features only; however, all models perform poorly with the highest accuracy of 86.2%, indicating that URL-based features alone are not adequate to detect phishing websites even with deep ANNs. Since language transformers learn to represent context-dependent text sequences, we hypothesize that they will be able to learn directly from the text in URLs to distinguish between legitimate and malicious websites. We apply three state-of-the-art deep transformers (BERT, ELECTRA, and RoBERTa) for phishing detection. Testing custom and standard vocabularies, we find that pre-trained transformers available for immediate use (with fine-tuning) outperform the model trained with the custom URL-based vocabulary. In addition, we test a thinner BERT transformer which is suitable for lightweight devices like mobiles, called MobileBERT. Our results emphasize that evaluation metrics of this model are competitive to other models in this study, yet the testing time is significantly less, making this model a choice for embedding phishing detection algorithms in mobile phones. Using pre-trained transformers to predict phishing websites from only URLs has five advantages: 1) requires little training time (230 to 320 s), 2) is more easily updatable than feature-based approaches because no pre-processing of URLs is required, 3) is safer to use because phishing websites can be predicted without physically visiting the malicious sites, 4) is easily deployable for real-time detection and is applicable to run on mobile devices, and 5) using a mobile specific transformer yields comparable performance and predicts 3 times faster than the other transformer models in this study.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.000 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.000 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.000 |
| Open science | 0.000 | 0.000 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from it