Commit
•
b6e1a5f
1
Parent(s):
b77c4b0
add corrected data analysis
Browse files
README.md
CHANGED
|
@@ -70,6 +70,37 @@ Here are two images showing the correlation coefficient and correlation of deter
|
|
| 70 |
|
| 71 |
|
| 72 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 73 |
For training, I split the classification task into two stages in anticipation of the limited availability of online phishing websites due to their short lifespan, as well as the possibility that research done on phishing is not up-to-date:
|
| 74 |
1. a small multilingual BERT model to output the confidence level of a URL being malicious to model #2, by finetuning on 2,436,727 legitimate and malicious URLs
|
| 75 |
2. (probably) LightGBM to analyze the confidence level, along with roughly 19 extracted features
|
|
|
|
| 70 |
|
| 71 |
|
| 72 |
|
| 73 |
+
Let's exmain the correlations one by one and cross out any unreasonable or insignificant correlations.
|
| 74 |
+
|
| 75 |
+
| Variable | Justification for Crossing Out |
|
| 76 |
+
|-----------------------------|------------------------------------- |
|
| 77 |
+
| ~~redirects~~ | contracdicts previous research (as redirects increase, is_malicious tends to decrease by a little) |
|
| 78 |
+
| ~~not_indexed_by_google~~ | 0.00 correlation |
|
| 79 |
+
| ~~email_submission~~ | contracdicts previous research |
|
| 80 |
+
| request_url_percentage | |
|
| 81 |
+
| ~~url_anchor_percentage~~ | contracdicts previous research |
|
| 82 |
+
| ~~meta_percentage~~ | 0.00 correlation |
|
| 83 |
+
| script_percentage | |
|
| 84 |
+
| link_percentage | |
|
| 85 |
+
| ~~mouseover_changes~~ | contracdicts previous research & 0.00 correlation |
|
| 86 |
+
| ~~right_clicked_disabled~~ | contracdicts previous research & 0.00 correlation |
|
| 87 |
+
| ~~popup_window_has_text_field~~ | contracdicts previous research |
|
| 88 |
+
| ~~use_iframe~~ | contracdicts previous research |
|
| 89 |
+
| ~~has_suspicious_ports~~ | contracdicts previous research |
|
| 90 |
+
| ~~external_favicons~~ | contracdicts previous research |
|
| 91 |
+
| TTL (Time to Live) | |
|
| 92 |
+
| ip_address_count | |
|
| 93 |
+
| ~~TXT_record~~ | all websites had a TXT record |
|
| 94 |
+
| ~~check_sfh~~ | contracdicts previous research |
|
| 95 |
+
| count_domain_occurances | |
|
| 96 |
+
| domain_registration_length | |
|
| 97 |
+
| abnormal_URL | |
|
| 98 |
+
| age_of_domain | |
|
| 99 |
+
| page_rank_decimal | |
|
| 100 |
+
|
| 101 |
+
|
| 102 |
+
## Pre-training Ideas
|
| 103 |
+
|
| 104 |
For training, I split the classification task into two stages in anticipation of the limited availability of online phishing websites due to their short lifespan, as well as the possibility that research done on phishing is not up-to-date:
|
| 105 |
1. a small multilingual BERT model to output the confidence level of a URL being malicious to model #2, by finetuning on 2,436,727 legitimate and malicious URLs
|
| 106 |
2. (probably) LightGBM to analyze the confidence level, along with roughly 19 extracted features
|