Inspiration
HumanZero is an AI detection platform that identifies AI-generated images and videos with high accuracy. You upload media, and we analyze it to determine whether it’s authentic or synthetic. We built this because AI content is becoming incredibly realistic. Very soon, humans won’t be able to reliably distinguish real from fake
How we built it
We built the models by using the pre-trained OpenAI CLIP ViT/L-14 model as a baseline, whose output layer was removed and then finetuned on a model trained on a custom dataset. For the image classification model, we made use of a multi-layer perceptron, while for the video classification, we relied on an LSTM. On the other hand, the website made use of a Flask backend and HTML and CSS frontend.
Challenges we ran into
Achieving a high accuracy and adequate generalization in the video models due to limited compute power and a short time-frame.
Accomplishments that we're proud of
We are proud of being able to achieve an industry-grade accuracy, which is resilient to a high degree of FGSM attacks.
What we learned
What's next for Human Zero
We hope that in the future, we are able to train detection models to also differentiate between real and AI-generated voice recordings.
Log in or sign up for Devpost to join the conversation.