Inspiration
As incoming first-year students at our respective universities, we recognized a pressing issue that significantly impacts the safety and security of our fellow peers. We've observed numerous instances of students falling victim to prevalent scams, including deceptive job or research position offers, fraudulent financial aid claims, and the inadvertent clicking of phishing links that aim to compromise personal information. The consequences are not just monetary; students often lose significant amounts of money and countless hours grappling with the aftermath of these all-too-common phishing tactics.
What it does
By analyzing the textual content and context of emails, our AI model assigns scores indicating the likelihood of an email being a phishing attempt. Users simply input the email text, and our AI swiftly evaluates it against a vast dataset of phishing and non-phishing emails.
How we built it
Our AI solution leverages the RobertaForSequenceClassification model which is a specialized variant of the BERT-cased model. We've fine-tuned this model using datasets filled with examples of phishing emails. In this process we used libraries and frameworks such as PyTorch, Pandas NumPy, Scikit-learn, Accelerate, and Datasets by Hugging Face. As for our User Interface, we first designed our web app using Figma before fully implementing it using JavaScript, HTML and CSS.
Challenges we ran into
One significant hurdle was our limited experience in Machine Learning, as we were venturing into new territory. We had to overcome the learning curve of understanding complex models and fine-tuning them for our specific task of email phishing detection. We also struggled in connecting our Front-end and Back-end in order to make the web-app functional with our AI tool. We approached this challenge by setting up a node server, and integrating AJAX for communication between our front-end and back-end but unfortunately we weren’t able to get this aspect fully working within the timeframe. Regardless, these challenges provided us with valuable learning experiences and ultimately strengthened our problem-solving skills.
Accomplishments that we're proud of
One of our major achievements is successfully building an AI-powered solution for email phishing detection, even though we had limited prior experience in Machine Learning. Furthermore, despite being a team of individuals who have never collaborated before, we effectively harnessed our collective expertise to create a unified solution. Navigating through the challenges of working together for the first time, we felt proud to have cultivated an environment of open communication and mutual respect, resulting in a harmonious and productive team dynamic.
What we learned
Through this project, we embarked on a steep learning curve that enriched our skillset and understanding in various domains. Our journey taught us the intricacies of Machine Learning, from preprocessing data to fine-tuning models. We gained hands-on experience in utilizing powerful libraries like Transformers and Hugging Face to create an effective phishing detection system.
What's next for Phish Net
Our immediate focus is on refining and streamlining the integration of our AI model into our web application by establishing robust communication between Node.js and AJAX. This will ensure seamless utilization of our AI for users directly through our platform. Furthermore, we plan to extend PhishNet's accessibility by developing a browser extension, making it even more convenient for users to safeguard against phishing threats while browsing.
Looking ahead, we aim to enhance the sophistication of our AI model. Beyond just determining phishing likelihood and providing a score, we envision our AI to offer granular insights. This means it will not only identify potentially malicious emails but also provide users with specific details about why certain parts of an email should be flagged. This added layer of analysis will empower users with deeper insights into potential threats, enabling them to make more informed decisions about their digital security.
Log in or sign up for Devpost to join the conversation.