Published Paper Details:

There are number of users who purchase products online and make payment through various websites. There are multiple websites who ask user to provide sensitive data such as username, password or credit card details etc. often for malicious reasons. This type of websites is known as phishing website. Phishing attacks are becoming more common and sophisticated, putting Internet users at risk. While these assaults have shown robust to a wide number of countermeasures presented by academia, business, and research groups, machine learning algorithms look to be a viable option for discriminating between phishing and authentic websites. Existing machine learning algorithms for phishing detection have three major drawbacks. The first issue is that there is neither a framework for extracting features and maintaining the dataset up to date, nor an updated collection of phishing and genuine websites. The second point of concern is the vast number of features employed, as well as the absence of supporting evidence for the characteristics used to train the machine learning classifier. The final point of concern is the sort of datasets utilised in the research, which appear to be unwittingly skewed in terms of URL or content-based attributes. The development of our open-source and extendable system to extract features and produce up-to-date phishing dataset is described in this thesis. We integrated 29 distinct characteristics into our framework, dubbed Fresh-Phish, to determine whether a particular website is authentic or phishing. We constructed a dataset of 6,000 websites with these qualities, 3,000 of which were malicious and 3,000 of which were legitimate, and evaluated our technique using 26 features published in previous work and three novel features. We concentrate on this aspect of phishing websites and design features that investigate the domain name's relationship to the website's key elements. Our study varies from the previous state-of-the-art in that our feature set assures that a dataset has low or no bias. On a sample dataset, our learning model achieves a true positive rate of 98 percent and a classification accuracy of 97 percent using only seven features. Our per data instance processing and classification is 4 times quicker for authentic websites and 10 times faster for phishing websites when compared to state-of-the-art work. We also show the drawbacks of utilising URL-based characteristics, since they are likely to be skewed towards dataset acquisition and consumption.