Automated Machine Learning Approaches for Phishing Detection and Prevention

The most common type of financial cybercrimes in the highly digitalized economy of India is also identified to be social engineering attacks/phishing. According to the Ministry of Home Affairs, in the financial year 2023-24, cybercriminals have withdrawn ₹22,845.73 Crores from citizens through online scams. It is also identified in the context of global cybercrimes that in the year 2023, the number of attacks recorded by the Anti-Phishing Working Group is 4.9 million. This is the highest number of cybercrimes ever recorded in history.The tools used for detection are mostly unimodal and reactive, utilizing 'black box' systems that are impossible to interpret. This project proposes PhishShield, which is a proactive multi-vector AI framework used for detecting cyber threats in real time, which is developed in the form of a publicly available Web Dashboard as well as a lightweight Chrome Browser Extension. This AI framework utilizes two highly efficient machine learning classifiers, namely, Naive Bayes Classifier with TF-IDF Vectorization for semantic markers in unstructured text data such as SMS and Email, as well as Decision Tree Classifier for structural features in URL data. The accuracy of the URL classifier is enhanced with the integration of Real-Time OSINT Heuristics. The output of the AI framework is analyzed, and the results are compared with the output of the Explainable AI module, which provides a definite two-class output in the form of 'Verified Safe' and 'Threat Detected' with Logic Trace. The testing of the AI framework is conducted, which provides promising results in the form of 92.6% accuracy with Naive Bayes NLP Classifier for detecting semantic threats as well as 90.7% accuracy with Decision Tree Classifier and OSINT Heuristics for detecting deceptive URLs.