Inspiration
Recently we have noticed an influx in elaborate spam calls, email, and texts. Although for a native English and technologically literate person in Canada, these phishing attempts are a mere inconvenience, to susceptible individuals falling for these attempts may result in heavy personal or financial loss. We aim to reduce this using our hack. We created PhishBlock to address the disparity in financial opportunities faced by minorities and vulnerable groups like the elderly, visually impaired, those with limited technological literacy, and ESL individuals. These groups are disproportionately targeted by financial scams. The PhishBlock app is specifically designed to help these individuals identify and avoid scams. By providing them with the tools to protect themselves, the app aims to level the playing field and reduce their risk of losing savings, ultimately giving them the same financial opportunities as others.
What it does
PhishBlock is a web application that leverages LLMs to parse and analyze email messages and recorded calls.
How we built it
We leveraged the following technologies to create a pipeline classify potentially malicious email from safe ones. Gmail API: Integrated reading a user’s email. Cloud Tech: Enabled voice recognition, data processing and training models. Google Cloud Enterprise (Vertex AI): Leveraged for secure cloud infrastructure. GPT: Employed for natural language understanding and generation. numPy, Pandas: Data collection and cleaning Scikit-learn: Applied for efficient model training
Challenges we ran into
None of our team members had worked with google’s authentication process and the gmail api, so much of saturday was devoted to hashing out technical difficulties with these things. On the AI side, data collection is an important step in training and fine tuning. Assuring the quality of the data was essential
Accomplishments that we're proud of
We are proud of coming together as a group and creating a demo to a project in such a short time frame
What we learned
The hackathon was just one day, but we realized we could get much more done than we initially intended. Our goal seemed tall when we planned it on Friday, but by Saturday night all the functionality we imagined had fallen into place. On the technical side, we didn’t use any frontend frameworks and built interactivity the classic way and it was incredibly challenging. However, we discovered a lot about what we’re capable of under extreme time pressures!
What's next for PhishBlock
We used closed source OpenAI API to fine tune a GPT 3.5 Model. This has obvious privacy concerns, but as a proof of concept it demonstrate the ability of LLMs to detect phishing attempts. With more computing power open source models can be used.
Log in or sign up for Devpost to join the conversation.