DeepSee

We were inspired since we are in a unique position where we are swarmed by mass content produced by generative AI but we are still able to reason about it and want to be able to pass this to other generations We built it using python and the pytorch libraries to train the AI model and gradCAM for the backend as well as utilising flask, JavaScript, HTML and CSS for the frontend We ran into challenges since we initially used the tensorflow library to train the CNN which worked but when we tried to run the gradCAM it was not working at all and came to the realisation that there was a problem with the library itself when implementing it which took a lot of time It was the first time many of us had coded in JavaScript and we were able to gain a deeper understanding of how interpretable AI works Our application is designed for use in a classroom environment, providing an interactive and educational experience for students. A teacher can sign up a group of students, who then compete in three different game modes to earn points. The first game serves as the primary learning tool, where students determine whether an image is AI-generated or real. If they incorrectly classify an AI-generated image as real, they are shown both the original image and a heatmap generated by Grad-CAM, highlighting the key areas that influenced the CNN’s decision. This visual feedback helps students understand the distinguishing features of AI-generated content. The other two game modes introduce a competitive element to reinforce learning. One is a rapid-fire timed challenge, and the other requires players to identify a specific number of real images from a set of six. These modes gamify the experience, making learning more engaging and enjoyable. What’s next for DeepSee? We would like to extend our learning environment to include more potential threats of misinformation, such as posts on social media or audio files using nlp and extending our use of ai to llms. We would also like to further refine our CNN and grad-CAM to improve its efficacy.

Built With

• flask
• javascript
• python
• pytorch