Created a tool that can take a url as an input and return whether the url is malicous, phishing, web defaced, or benign.

Feature Engineered a single column data set into multiple additional columns and feaures based on url stats

Trained several models for classification to include Logistic, Decision Tree, Random Forest, and K Nearest Neighbors.

Built client facing API using Flask

Python Version: 3.10

Packages: pandas, numpy, seaborn, matplotlib, sklearn, flask, jsonify, json, pickle, requests

Dataset: https://www.kaggle.com/datasets/sid321axn/malicious-urls-dataset

Upon seeing the data, I identified that it was not enough information for a classification model to make reasonable predictions. I added several new columns based on domain knowledge and statistics.

length of url

how many parts of seperated by "."

length of suffice (".com",".ru",".md", etc.)

ip address present

file extension present

shortened links present ("bit.ly", "t.co", tinyurl.com", etc.)

took a list of special characters and got the count of each in a url and created columns of the results

identified distribution and correlation between columns and the classification value, Used groubyby, scatterplots, and heat maps for visualizations. Highlights Below

Identified that the benign url were almost all over the place, but you can see that of the malware, phishing, and defacement url were more tightly clustered.

Malware url, had a significantly higher chance to have an ip address present.

The dataset is very imbalanced , with benign urls having over 50% of the values in the dataset, depending on the results of the models, I would have considered merging the data between the other columns and creating a binary classification vs. a multiclass

For Model Building I knew a couple models that I wanted to try ahead of time that I usually get good results from, First I split the data in a 70/30 train test split. I then dropped the "type" column that we are trying to predict from my X data and then dropped the "url" column , since I already have the important features in the new engineered columns. I then used Sklearn's StandardScaler to scale all of the X data. Created a function that takes in all of the takes and trained them all and compared results based on accuracy, recall, precision and f1 scores. Depending on the results of the training I would have used different hyperparameters for each model via a gridsearch.

Both Tree Based Models performed the best, but the RandomForestClassifier was the best overall

Accuracy: 93.89999999999999%

Precision: 91.8%

Recall: 91.3%

F1-Score: 91.5%

Accuracy: 78.0%

Precision: 62.1%

Recall: 56.699999999999996%

F1-Score: 58.199999999999996%

Accuracy: 94.5%

Precision: 92.9%

Recall: 92.0%

F1-Score: 92.4%

Accuracy: 91.8%

Precision: 89.5%

Recall: 87.7%

F1-Score: 88.5%

Confusion matrix of the True Labels and Predicted Label,

I then pickled the model and created a flask API endpoint hosted on a local server that takes in a single url and identifies the class