Investigating the Influence of Feature Sources for Malicious Website Detection
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
Malicious websites in general, and phishing websites in particular, attempt to mimic legitimate websites in order to trick users into trusting them. These websites, often a primary method for credential collection, pose a severe threat to large enterprises. Credential collection enables malicious actors to infiltrate enterprise systems without triggering the usual alarms. Therefore, there is a vital need to gain deep insights into the statistical features of these websites that enable Machine Learning (ML) models to classify them from their benign counterparts. Our objective in this paper is to provide this necessary investigation, more specifically, our contribution is to observe and evaluate combinations of feature sources that have not been studied in the existing literature—primarily involving embeddings extracted with Transformer-type neural networks. The second contribution is a new dataset for this problem, GAWAIN, constructed in a way that offers other researchers not only access to data, but our whole data acquisition and processing pipeline. The experiments on our new GAWAIN dataset show that the classification problem is much harder than reported in other studies—we are able to obtain around 84% in terms of test accuracy. For individual feature contributions, the most relevant ones are coming from URL embeddings, indicating that this additional step in the processing pipeline is needed in order to improve predictions. A surprising outcome of the investigation is lack of content-related features (HTML, JavaScript) from the top-10 list. When comparing the prediction outcomes between models trained on commonly used features in the literature versus embedding-related features, the gain with embeddings is slightly above 1% in terms of test accuracy. However, we argue that even this somewhat small increase can play a significant role in detecting malicious websites, and thus these types of feature categories are worth investigating further.
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 | 0.001 |
| Science and technology studies | 0.002 | 0.000 |
| Scholarly communication | 0.000 | 0.000 |
| Open science | 0.001 | 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