EKILA: Synthetic Media Provenance and Attribution for Generative Art

Kar Balan, Shruti Agarwal, Simon Jenni, Andy Parsons, Andrew Gilbert, John P. Collomosse

Abstract

We present EKILA; a decentralized framework that enables creatives to receive recognition and reward for their contributions to generative AI (GenAI). EKILA proposes a robust visual attribution technique and combines this with an emerging content provenance standard (C2PA) to address the problem of synthetic image provenance – determining the generative model and training data responsible for an AI-generated image. Furthermore, EKILA extends the non-fungible token (NFT) ecosystem to introduce a tokenized representation for rights, enabling a triangular relationship between the asset’s Ownership, Rights, and Attribution (ORA). Leveraging the ORA relationship enables creators to express agency over training consent and, through our attribution model, to receive apportioned credit, including royalty payments for the use of their assets in GenAI.

Introduction

Generative AI models, such as Stable Diffusion and DALL-E, are trained on vast datasets often without the explicit consent or compensation of the original creators. This lack of transparency creates a "black box" where identifying the specific training images that influenced a synthetic output is nearly impossible. This paper, a collaboration between the University of Surrey and Adobe Inc., introduces EKILA, a decentralized framework designed to track the "ancestry" of AI-generated images and ensure creators receive recognition and reward for their contributions.

The ORA Framework

EKILA addresses the shortcomings of current NFT implementations, which typically track ownership but fail to manage usage rights or provenance. The framework is built upon the ORA Triangle:

  • Ownership (O): Managed via standard NFTs on a distributed ledger.
  • Rights (R): A novel "Rights Contract" (smart contract) that allows creators to issue tokenized licenses specifically for AI training.
  • Attribution (A): Utilizing the C2PA (Coalition for Content Provenance and Authenticity) standard to embed manifests that list the training data and generative model used.

Visual Attribution and Apportionment

A core technical contribution of EKILA is its Visual Attribution and Apportionment model. Rather than analyzing images globally, the system employs a patch-based approach to identify specific visual "memorization"—instances where a model has replicated textures or objects from its training set.

EKILA's attribution pipeline: identifying a synthetic image's 'ancestry' by matching patches to specific training data.

The process involves several key steps:

  1. Patchified Fingerprinting: A ResNet-50 encoder identifies similar patches within a large-scale database.
  2. Match Verification: A three-layer Multi-Layer Perceptron (MLP) filters out false positives by comparing feature maps.
  3. Credit Apportionment: A mathematical formula weighs the contribution of each identified training image, enabling fair royalty distribution among multiple artists.

Experimental Evaluation

The researchers evaluated EKILA using several large-scale datasets, including a 10M image subset of LAION-400M, Adobe's IPF-Stock (200M assets), and the style-sensitive BAM-FG dataset.

Key results include:

  • Superior Accuracy: EKILA achieved an 83.33% Recall@1 for image attribution on the LAION dataset, significantly outperforming ViT-CLIP (56.25%).
  • Robust Verification: The Match Verification Model achieved an AUC of 0.740, surpassing simple embedding distance and perceptual metrics like LPIPS.
  • Scalability: The system proved effective at searching and attributing images across millions of assets, demonstrating real-world viability for creative economies.

Discussion and Future Work

EKILA introduces the concept of Rights as Decoupled Tokens, allowing artists to retain ownership of an image while separately licensing its training rights. By linking ownership, rights, and attribution, the system also mitigates "copy-minting" by providing a verifiable C2PA manifest of an image's creation history.

While the model effectively identifies visual similarity as a proxy for attribution, the authors note that proving a direct causal link remains a challenge. Future work will focus on maturing the legal frameworks and digital rights languages required to make such a decentralized creative economy sustainable.

Summary updated

Abstract

We present EKILA; a decentralized framework that enables creatives to receive recognition and reward for their contributions to generative AI (GenAI). EKILA proposes a robust visual attribution technique and combines this with an emerging content provenance standard (C2PA) to address the problem of synthetic image provenance – determining the generative model and training data responsible for an AI-generated image. Furthermore, EKILA extends the non-fungible token (NFT) ecosystem to introduce a tokenized representation for rights, enabling a triangular relationship between the asset’s Ownership, Rights, and Attribution (ORA). Leveraging the ORA relationship enables creators to express agency over training consent and, through our attribution model, to receive apportioned credit, including royalty payments for the use of their assets in GenAI.
BibTeX
@misc{balan2023ekila,
  author = {Kar Balan and Shruti Agarwal and Simon Jenni and Andy Parsons and Andrew Gilbert and John P. Collomosse},
  title = {{EKILA: Synthetic Media Provenance and Attribution for Generative Art}},
  year = {2023},
  eprint = {2304.04639},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2304.04639},
}
Figure