Our team was motivated to take on this opportunity by our personal interests in machine learning, the integration of art and technology, and fascination by AI-generated artwork. We were inspired by the rapid growth of ChatGPT, and recent news of attempts to build softwares that can distinguish between writing and work by humans and ChatGPT.
What it does
Our current code takes in a dataset of images of artwork that are known to be either real or AI-generated (synthetic). It then processes each image into the same format and size, and labels them as 0 for real or 1 for synthetic, and stores the real images, synthetic images, and labels as separate numpy arrays to work with. It then uses 80% of the images as a training set to learn from, and 20% as a testing set, using linear regression classification, and outputs the % accuracy of the results from the testing set, which we were able to get over 70%.
How we built it
We used Python for the back end code, and imported LinearRegression from scikit-learn, an existing machine learning library, in the implementation of our solution to build a working prototype. We also used WIX to design our idea for the front end website that would eventually connect with the code.
Challenges we ran into
A challenge we ran into was figuring out how to actually implement the image processing and classification systems once we came up with the idea. As first year students with very little experience with machine learning, we consulted many open online sources and mentors, and also made use of an existing classification model, logarithmic regression, in our code.
Accomplishments that we're proud of
We are extremely proud of completing a working prototype and reaching an accuracy of over 70% with our test sample dataset.
What we learned
As it was most of the team's first time working with machine learning, almost everything was new to us and we were so happy to learn how to build machine learning models. We especially learned lots about CNN classification models and image processing.
What's next for AIdentify
Our next steps are to improve the accuracy of the classification system by training it with larger, more diverse datasets, and also finding ways to refine and optimize the algorithm. We would also like to seamlessly integrate the back end code with the front end website for public users. Ideally, users would be able to use an easy-to-navigate website to upload any image of a piece of art, and have it return an accurate classification of whether it is real or AI-generated.
Log in or sign up for Devpost to join the conversation.