Inspiration
I first thought of ProtectPII when my computer science teacher shared how often people accidentally shared personally identifiable information(PII) through social media posts. During the lesson, she shared examples of posts that could reveal PII (a picture of your first car that displays the license plate or a photo from your birthday party that reveals your date of birth). I was surprised that I hadn’t even thought of the possible risks of sharing photos of some of the scenarios she mentioned. As a result, I created ProtectPII, which alerts users to risks associated with social media posts that they may not have even considered.
What it does
Users input the picture and the caption they plan to post into ProtectPII, which alerts them of which category of risk the PII in their photo places under.
How we built it
I used OpenAI’s CLIP to produce descriptions for the user’s image. I then created an array of potential scenarios where PII might be exposed and gave each scenario a risk level. I then compared CLIP’s description to all of the scenarios listed and compared them. If the similarity was above a certain threshold, the image would be marked as containing PII. I then used the risk level associated with each scenario to sort the images into three categories of risk: safe, confidential, and high risk.
Challenges we ran into
In earlier tests, ProtectPII would flag a safe image(pictures of scenery or animals) as confidential or high risk. I had to adjust the threshold for similarity to ensure images were not incorrectly marked.
What's next for ProtectPII
I wish to make ProtectPII more sensitive to more subtle forms of PII that might indicate to malicious people information like what school you go to or when you are on a vacation. I also wish to add a feature that pinpoints exactly what makes the post a risk, so users do not have to worry about trying to identify what part of their post is being flagged.
Log in or sign up for Devpost to join the conversation.