Overlai White Paper

Abstract

​

AI protection is a growing concern, and Overlai addresses this challenge through a combination of innovative techniques. Starting with blockchain to establish immutable records of asset ownership and approved usage, our solution employs various technologies to prevent AI scraping, ensuring the integrity and protection of authentic digital assets.

​

1 Introduction

​

The advent of text-to-image generator models has undeniably reshaped the landscape of the digital art industry. Models like StableDiffusion or MidJourney allow users to effortlessly generate intricate, high-resolution artwork from simple text prompts.

 

Such AI-generated art has not only received accolades at established art conventions but has also adorned magazine covers, illustrated children's books, and even found its way into video games.

 

These powerful models, catalyzed by VC funding and continuous training on vast datasets scraped from online art repositories, are gaining rapid popularity. Platforms like MidJourney boast millions of users, generating hundreds of thousands of AI art images daily. However, the impact of these models on independent artists is significant.

 

Mimicry attacks, where models replicate specific artistic styles, jeopardize the livelihoods of countless independent artists globally. The consequences range from the devaluation of artistic training to the displacement of original art in search results, disrupting artists' ability to showcase their work to potential customers.

​

In response to these challenges, we introduce Overlai, a comprehensive solution designed to protect digital assets from AI mimicry. Leveraging blockchain and cutting-edge technologies, Overlai ensures immutable records of asset ownership and approved usage, offering multi-tiered protection against AI scraping.

​

1.1 Motivation

​

The detrimental effects of AI models on the art community are evident. Artists are facing challenges from legal and regulatory perspectives, often resorting to online activism, lawsuits, and petitions. However, these measures are time-consuming and challenging to enforce internationally. Overlai seeks to provide a technical alternative, offering a robust defense against AI mimicry. Our system introduces advanced techniques, including blockchain-based NFTs, decentralized storage on IPFS, EXIF data embedding, steganography, and the ability for creators to have provenance over their creative works.

​

1.2 Contributions

​

Our work with Overlai encompasses the following key contributions:

​

• Engagement with top professionals and the art community, with user studies to comprehend their views and concerns regarding AI art's impact on their careers and community.

• Introduction of Overlai, a system employing minimal perturbations to mislead AI models and protect artists from style mimicry. Surveyed artists find the perturbations subtle enough not to disrupt the value of their art.

• Validation of Overlai's effectiveness through user studies, demonstrating high success rates in disrupting style mimicry, even in challenging scenarios.

​

In subsequent sections, we delve into the design, implementation, and evaluation of Overlai, emphasizing its robustness against countermeasures and discussing deployment experiences.

​

​

2 Blockchain

​

The blockchain mechanism of Overlai consists of two main components: the NFT minter and the metadata/image storage. Both of these components are key in establishing our immutable record of ownership and approved usage.

​

2.1 NFT

​

Each photo uploaded on Overlai is minted as an NFT using the ERC-721 Non-Fungible Token Standard. All NFTs have a uint256 variable called tokenId, so for any ERC-721 Contract, the pair contract address, uint256 tokenId must be globally unique.

​

2.2 Decentralized Storage

​

Each image is minted to IPFS where it is appended with EXIF data, watermarked and then assigned the Rights Managed Contract.

​

• IPFS (InterPlanetary File System) is a protocol, hypermedia and file sharing peer-to-peer network for storing and sharing data in a distributed file system. IPFS uses content-addressing to uniquely identify each file in a global namespace connecting IPFS hosts.

​

• Rights Managed Contract - The metadata of each NFT includes a custom Rights Managed Contract, giving artists an immutable record of their asset history and acting as a legal contract, stating the restrictions around AI usage.

​

• Copyright information is appended into the EXIF metadata of the image, assigning ownership of the underlying asset.

​

• EXIF data is assigned to the NFT Metadata. Exchangeable Image File Format (officially Exif, according to JEIDA/JEITA/CIPA specifications) is a standard that specifies formats for images, sound, and ancillary tags used by digital cameras (including smartphones), scanners and other systems handling image and sound files recorded by digital cameras. The specification uses the following existing encoding formats with the addition of specific metadata tags: JPEG lossy coding for compressed image files, TIFF Rev. 6.0 (RGB or YCbCr) for uncompressed image files, and RIFF WAV for audio files (linear PCM or ITU-T G.711 μ-law PCM for uncompressed audio data, and IMA-ADPCM for compressed audio data). It does not support JPEG 2000 or GIF encoded images.

 

3 Steganography

​

Steganography, the art of concealing information within other data, plays a crucial role in Overlai's multi-tiered protection strategy. With our innovative approach, we embed a unique identifier for each image. This identifier is machine-readable, ensuring fast processing by dataset processors, yet remains invisible to human observers. This not only expedites the identification of protected images but also ensures compatibility with popular AI datasets and withstands compression on widely used social media platforms.

​

3.1 Technique

​

The technique employed by Overlai involves pre-cropped JPEG encoding combined with robust watermarks strategically embedded in the Discrete Cosine Transform (DCT) coefficients. This approach ensures the integrity of Overlai watermarks, optimizing them for dataset Watermark Detectors. Our method employs a carefully crafted identifier that withstands compression, making it resilient even in environments where images undergo significant data reduction, such as those encountered on social media platforms.

​

3.2 Poison Layer

​

As part of our continuous commitment to enhancing protection, Overlai is actively exploring the development of a hybrid poisoning model. This model aims to serve as an additional defense layer, further fortifying Overlai's capabilities in safeguarding digital assets against unauthorized usage. The poison model will be designed to complement existing protection mechanisms, exploring innovative ways to counter evolving threats in the realm of AI mimicry and digital asset misuse.

​

4 Opt-Outs

​

Overlai takes a proactive approach to dataset monitoring through a feature we call "Opt-Outs." Leveraging forks to the GitHub repository of dataset watermark detectors, our system establishes communication channels with datasets, flagging Overlai watermarks for blocking.

​

4.1 Mechanism

​

Instead of traditional methods involving the scanning of datasets for potential image scraping, Overlai's Opt-Outs feature focuses on the detection of watermark usage. This strategic shift allows for more cost-effective and efficient monitoring of datasets.

​

4.2 Dataset Communication

​

Overlai actively communicates with datasets by integrating with their watermark detectors. When Overlai watermarks are detected, the system promptly signals the dataset to take appropriate measures, such as blocking or excluding the flagged content.

​

4.3 Advantages

​

The Opt-Out mechanism provides several advantages:

​

• Cost-Effective Monitoring - By specifically targeting watermark usage, Overlai streamlines the monitoring process, making it a more cost-effective solution for dataset owners.

• Timely Intervention - Rapid communication with datasets enables timely intervention, preventing the proliferation of protected content within unauthorized repositories.

• Targeted Enforcement - Rather than broad scanning for image scraping, Overlai's approach focuses on the specific identification and blocking of watermarked content, ensuring targeted enforcement. In summary, the Opt-Out feature is a pivotal component of Overlai's strategy, reinforcing its commitment to proactive and efficient protection against unauthorized use of digital assets.

​

5 Certificate and Copyright

​

In the realm of digital asset protection, Overlai goes beyond prevention and proactively produces a certificate for each processed image. These certificates serve as a crucial last line of defense, ready to enforce a takedown or cease and desist in the rare event that our front-facing protections are breached.

​

5.1 Certificate Details

​

For every processed picture, Overlai generates a comprehensive certificate, including:

​

• Links to IPFS - Each certificate provides links to the InterPlanetary File System (IPFS), ensuring decentralized and tamper-resistant storage of the image.

• NFT Contract Address/ID - The certificate includes the Non-Fungible Token (NFT) contract address and ID, reinforcing the immutable record of ownership on the blockchain.

• Image File - A copy of the image is embedded in the certificate, allowing for quick verification and comparison.

• EXIF Data - Overlai ensures that the EXIF metadata is included in the certificate, preserving information about the image and its origin.

• Date/Time of Processing - The certificate contains a timestamp, documenting when the image was processed by Overlai.

​

5.2 Last Line of Defense

​

This certificate serves as an additional layer of security, ready to be invoked in case of a breach. If our primary protections are compromised, the certificate becomes instrumental in initiating legal actions, such as takedown requests or cease and desist orders.

 

By linking crucial information directly to the blockchain, Overlai establishes a robust and tamper-proof record that strengthens the position of artists and content creators. In essence, Overlai's commitment to blockchain technology ensures that the protection journey begins and ends on a secure foundation.

​

​

Overlai White Paper Disclaimer:

​

This white paper is intended to provide insights into the technology and features of Overlai. While every effort has been made to ensure the accuracy and reliability of the information presented herein, Overlai makes no representations or warranties, express or implied, regarding the completeness, reliability, or suitability of the information for any particular purpose.

​

Readers are advised to conduct their own due diligence and seek professional advice before relying on the information contained in this white paper. Overlai shall not be held liable for any direct, indirect, incidental, consequential, or punitive damages arising from the use of or reliance on this white paper or its contents.

​

The contents of this white paper are subject to change without notice, and Overlai reserves the right to update, modify, or discontinue any aspect of the platform described herein at any time.

​

​

References

​

[1] Carras, Christi. “These entertainment jobs are most vulnerable to AI, study says.” Los Angeles Times, https://www.latimes.com/entertainment-arts/business/story/2024-01-30/ai-artificial-intelligence-impact-report-entertain ment-industry. Accessed 31 January 2024.

​

[2] NOVECK, JOCELYN, and MATT O'BRIEN. “Visual artists fight back against AI companies for repurposing their work.” AP News, 31 August 2023, https://apnews.com/article/ai-technology-arts-and-entertainment-business-49cf7846b6e0ffce72fa2e7a 379025e1. Accessed 31 January 2024.

​

[3] “ERC-721 Non-Fungible Token Standard | ethereum.org.” Ethereum, 19 November 2023, https://ethereum.org/developers/docs/standards/tokens/erc-721. Accessed 31 January 2024.

​

[4] IPFS: An open system to manage data without a central server, https://ipfs.tech/. Accessed 31 January 2024.

​

​