najmulhasan/PhishLlama_LoRA

PhishLlama LoRA: A LoRA Fine-Tuned Model for Phishing URL Detection

Model Details

• Model Name: PhishLlama LoRA
• Developed by: Najmul Hasan
• Model Type: LoRA fine-tuned version of meta-llama/Llama-3.2-1B
• Language: English
• License: Apache 2.0
• Base Model: meta-llama/Llama-3.2-1B
• Pipeline Tag: Text Classification
• Library: Transformers
• Dataset Used: PhiUSIIL Phishing URL (Website) - Reduced Dataset
• Training Data Reduction: The dataset was balanced and reduced to 50,000 samples while maintaining class distribution.

Model Description

PhishLlama LoRA is a LoRA fine-tuned version of Llama-3.2-1B, designed for phishing URL detection. It classifies URLs as either legitimate (1) or phishing (0).
The model was trained using a reduced version of the PhiUSIIL Phishing URL dataset to improve training speed, computational efficiency, and maintain predictive performance.

Performance

• Accuracy: 59%
• Macro F1-score: 0.58

Confusion Matrix:

Prediction / Actual	Phishing (0)	Legitimate (1)
Phishing (0)	3668 (True Negative)	2131 (False Negative)
Legitimate (1)	1954 (False Positive)	2247 (True Positive)

• True Positive (TP): 2247 (Legitimate detected as Legitimate)
• False Positive (FP): 1954 (Phishing misclassified as Legitimate)
• True Negative (TN): 3668 (Phishing detected correctly)
• False Negative (FN): 2131 (Legitimate misclassified as Phishing)

---