This is the sixth installment of our Tech article series. The goal of this multipart series is to discuss provenance at a high level and provide more technically inclined readers a full dive into how to execute digital provenance. You can find the full list of this series here :
Part 1: What is Provenance and why we need it
Part 2: What is C2PA, Add C2PA easily with Capture
Part 3: Take True Photo with C2PA Watermark from Mobile Phones
Part 4 : Add C2PA easily with Capture API
Part 5 : C2PA for Music and Video Platforms with Capture API
Part 6 : You are here
Overview
OpenAI’s recent C2PA and DALL-E 3 integration, shortly followed by TikTok represents a growing prioritization for content provenance. The Coalition for Content Provenance and Authenticity (C2PA) is an open standard that advocates for embedding metadata in digital media to verify its origin and other related information. As a leading player in the generative AI space, OpenAI's adoption of these standards underscores a shift towards responsible AI and sets a precedent for other AI platforms.
Since its introduction, many have expressed concerns surrounding the blurring boundaries between human and AI creations. Strong push-back from creator communities stem from their works being used without their consent in AI training models. All of these issues can be managed if there was a robust way of verifying digital content. One of the solutions? C2PA + Capture.
Integrating C2PA into existing systems can be challenging due to technical aspects and resource requirements. However, Capture powered by Numbers Protocol provides an accessible, easy-to-use toolkit designed to overcome these challenges. This solution allows generative AI companies to seamlessly integrate C2PA standards without significant additional investments in resources or infrastructure.
Contents
1. No-Code Generative AI Recap
2. Label AI-Generated Content
3. Acquiring the C2PA Injected Generative AI Image
4. Verifying C2PA contents on the C2PA verification site
This “How To” leverages Instill AI’s platform and Numbers Protocol’s Capture solution to build a traceable generative AI pipeline. For details on creating necessary accounts and acquiring Capture Authorization Token, please refer to the following tutorial videos.
Acquire Capture Authorization Token from the Dashboard.
Acquire Capture Authorization Token from API.
No-Code Generative AI Recap
Instill AI is a no/low code tool that allows us to easily access and create generative AI pipelines using different AI models such as Stable Diffusion or the latest Llama 3 released by Meta. Numbers Protocol’s Blockchain Provenance solution is integrated as the sole provenance connector. To construct a traceable text-to-image generative AI pipeline we will navigate to Instill’s pipeline constructor and build the pipeline in the following manner:
Configure your pipeline like shown above. Text-to-Image takes in text input which is fed into Stability AI to output an Image. This Image is stored along with metadata inputs onto the Numbers blockchain. As a final output we are interested in seeing our final image and the metadata stored on-chain.
You can refer to the following tutorial video for a full step-by-step breakdown of how to construct this pipeline. While this tutorial opted to use the Instill AI Model, we can easily select another AI model from the AI connector list.
Tutorial Video on creating fully functional and traceable generative AI with no-code Tools.
Labeling AI-generated content
Instill AI pipeline builder doubles as an interactive playground allowing us to test the AI pipeline without the need for API requests. Simply fill in the fields in the start operator and click “Run”.
Successful text-to-image creation and image registration to Numbers Protocol blockchain. Registered images with blockchain records are called “Asset”. Output includes final_image (see image) and asset-profile link containing on-chain records.
This process effectively serves as an AI watermark, ensuring that the generated content can always be traced back to its origin.
Acquiring the C2PA Injected Generative AI Image
You can acquire the registered C2PA injected image by using either Capture API or, Capture Dashboard. Of course, we also provide a video tutorial to help you on both methods.
Tutorial on how to label your registered image with C2PA.
Having successfully registered the generative AI image to Numbers Blockchain, we can acquire the C2PA injected file by using the Capture API, specifically the Download asset with C2PA injection API action. Using your preferred API platform or sending a curl request we can send the following POST request:POST
https://api.numbersprotocol.io/api/v3/assets/{nid}/c2pa/
For this API call we will need to populate our Header with our Authorization Token:Authorization Token <Capture Token>
Replacing the {nid} parameter with the nid url parameter in the asset profile link from the previous step we get the following:
Navigating to the url we can see the C2PA injected image.
This image has an invisible C2PA watermark labelled on it
Verifying C2PA contents on the C2PA verification site
The beauty of C2PA is the image itself appears unchanged, but in reality embedded in the media itself is metadata. At the same time it is challenging to tell if the image we just downloaded is a C2PA image. Luckily, C2PA has provided a verification site we can leverage to check our images contents.
By navigating to https://contentcredentials.org/verify we can upload our C2PA image and verify its contents.
The left panel contains a summary of the embedded information. See below for a full view of the embedded C2PA contents:
Verifying C2PA-lost AI content?
As OpenAI stated: “Metadata like C2PA is not a silver bullet to address issues of provenance. It can easily be removed either accidentally or intentionally.” Embedding C2PA metadata into digital media is a great way to make assets verifiable, but the content may be altered, compressed or cropped while sharing on social media. To address this challenge, Numbers Verify Engine can help match the content to the registered pool. The AI engine behind will find the matched content to proceed the verification. We will cover this in another article.
Final Remarks
C2PA is an attractive content provenance solution for generative AI and digital media that has been adopted by OpenAI. In addition to C2PA, Capture also adheres to the IPTC standard that ensures every content is easily searchable, shareable, and verifiable across different platforms and systems. Given the recent changes in AI regulation, which require that all generated AI images be labeled or watermarked, and considering that technological advancement often outpaces regulation, we aim to assist generative AI companies with the necessary tools suitable for compliance.For further consultation or to learn more about our solution, feel free to reach out to us.