DeepShield

Inspiration

We built DeepShield in response to a growing problem: it is getting harder and harder to tell whether a photo is real or not. We focused on giving people a way to protect legitimate photos at the moment they are created or shared through a digital signature.

What it does

In the Sign section, a user uploads an image and the app converts it into a PNG file. It then creates a log based on the image's exact pixel data, dimensions, and channels. That log is packaged into a signature payload with a timestamp and protected with an signature. The signed payload is embedded directly into the PNG metadata for hidden signing, and the user can download the signed file as a protected master copy.

In the Check section, the user uploads an image for verification. The app first looks for the embedded signature in the PNG metadata. If a signature is found, it validates the signature and recomputes the image log to confirm whether the pixels still match the original signed image. If both checks pass, the image is returned to the user as authentic. If the signature exists but the pixels do not match, the image is flagged as altered or tampered with.

If the uploaded image has no valid signature, DeepShield uses an image detection provider to estimate whether the image is likely authentic, likely AI-generated, or inconclusive. This gives users a practical fallback when they are checking unsigned images.

How we built it

We developed DeepShield as a comprehensive web application utilizing Next.js, TypeScript, and Tailwind CSS. The frontend features signing and verification interfaces with drag-and-drop uploads, real-time previews, and AI-generated probabilities.

We utilized Next.js API routes on the Node.js runtime for the backend. To sign the PNGs, we secure the payload with an HMAC-SHA256 signature utilizing a server-side secret that we create.

Challenges we ran into

A big challenged we faced was integrating data directly into PNG metadata while still being able to preserve the file format. Another challenge we faced was fine tuning our model for it to be accurate.

Accomplishments that we're proud of

We are proud in the fact that DeepShield goes beyond just identifying AI-generated images. It also provides users with a method to safeguard their photos by integrating a verification record straight into the image that they upload.

What we learned

We learnt that creating these safety tools requires both performing well technically from the model and an appealing user experience. Achieving a correct result is not enough if the user cannot comprehend what the system is presenting to them.