AutoAuth

An AI-powered agent that aims to protect both car insurance companies and buyers from visual fraud in vehicle images. Detects AI generated/tampered elements of an image and returns a confidence score.

Comment

Inspiration

The inspiration for this project came to me from an article I had come across some time back. It talked about the growing risk of AI-generated tampering in the world of car insurance. With the development of these image generation models, there is a growing concern that they will be used in insurance claim fraud cases.

What it does

How we built it

Challenges we ran into

The biggest challenge I ran into was with regards to training the model itself and figuring out the best/most effective way to implement the idea within the 10 hour window. There were no datasets online that pertained to my needs, different view of vehicles - generated by AI. Because of this, I had to write a script to create a bunch of these AI-generated images and handpick the ones I wanted to use to train my model. This was a long and lengthy process that required me to not only generate these images but then filter them and process them to ensure they matched the format of the non-AI generated images in order to yield a better model.

Accomplishments that we're proud of

// these part did not save as I forgot to click save and continue, responses in comments.

What we learned

What's next for AutoAuth

I believe that there is plenty of room for this to grow.

Built With

  • tensor
Share this project:

Updates

posted an update —

The responses to the questions did not save:

What it does It allows the user to input an image of a car from an angle and gives them a verdict on weather or not it was AI-generated or a real photo of a car. It also gives a natural language explanation of the decision based off of the verdict and confidence score.

How we built it It is built entirely in python and utilizes gradio for the frontend/UI. I trained the ResNet18 model from the pytorch on my datasets and then deployed it on the Tiber cloud. After the user uploads an image, the agent has a pipeline that processes the image and runs it through the model. It first removes the background and then converts it into a tensor. After the model prediction, we get a confidence score and verdict(targe/non-target).

Accomplishments that we're proud of I am most proud of the model training process as it required me to generate my own dataset and process it thoroughly to produce an accurate model. I am also proud of the natural language implementation as I believe it gave it a bit of a unique twist.

What we learned: I'd never used hugging face prior to this hackathon and didn't realize how useful it was. I'm glad that I got to learn about it and will definitely use it in the future.

What's next for AutoAuth Given more time, I would have liked to train separate models for the front, back and side views of the cars. Additionally I think some sort of heatmap implementation with Grad Cam would be useful as it would aid in decision explainability.

Log in or sign up for Devpost to join the conversation.