Phishing-website-detection-using-machine-learning

The project aims to build and evaluate machine learning models that can effectively classify websites as phishing or legitimate based on their features. The use of data balancing, visualization, and various models demonstrates a comprehensive approach to tackling the phishing detection problem.

---

Data Loading and Exploration

1. Dataset Loading: The project begins by loading the phishing.csv dataset.
2. Exploratory Analysis:
   • Used functions like head(), info(), and isnull().sum() to understand the data structure and check for missing values.

---

Data Balancing

• The target variable (class), which indicates whether a website is phishing or legitimate, was found to be imbalanced.
• To address this, the RandomOverSampler technique was employed, ensuring equal representation of both classes for better model training.

---

Data Visualization

Several visualizations were created to gain insights into the data:

1. Bar Plots:

   • Distribution of the target variable before and after balancing.
   • Comparison of counts for LongURL and ShortURL features.

2. Feature Importance:

   • Used ExtraTreesClassifier to visualize the most influential features.

3. Feature Relationships:

   • Explored the relationship between AnchorURL and HTTPS using a stacked bar chart.

4. Correlation Heatmap:

   • Displayed relationships between different features to identify patterns.

---

Model Training and Evaluation

Various machine learning models were trained to detect phishing websites:

• Random Forest
• Decision Tree
• SVM (Support Vector Machine)
• Naive Bayes
• Gradient Boosting
• LightGBM
• CatBoost

Evaluation Metrics:

• Accuracy scores and classification reports were generated for each model to assess their performance.

---

Model Comparison

• A bar plot was used to compare the accuracies of different models.
• This helped identify the best-performing model(s).

---

Hybrid Model Exploration

• The project explored hybrid models by combining different base models using VotingClassifier.
• The accuracies of these hybrid models were compared to assess improvements in performance.

---

Thank you 😊