Introduction
The rapid proliferation of Generative AI (GenAI) has fundamentally disrupted the landscape of digital intellectual property. While models like Stable Diffusion and DALL-E demonstrate remarkable creative capabilities, they are built upon billions of images often harvested without the explicit consent of, or compensation for, the original creators. This "Wild West" environment has left artists without a mechanism to track the usage of their work in synthetic media.
EKILA addresses this gap by introducing a decentralized framework that links AI-generated outputs back to their specific training data. By integrating robust visual attribution algorithms with blockchain-based smart contracts and the C2PA provenance standard, EKILA establishes a transparent royalty system where artists are automatically recognized and compensated when their work influences new synthetic creations.
The ORA Triangle: A New Model for Digital Rights
To move beyond the limitations of current Non-Fungible Tokens (NFTs)—which often lack mechanisms for tracking usage rights—the authors propose the ORA Triangle. This conceptual model manages three critical components of digital assets:
- Ownership (O): Tracked via standard ERC-721 NFTs to establish the current holder of the asset.
- Rights (R): A novel smart contract that issues tokenized licenses specifically for GenAI training and reuse.
- Attribution (A): The use of C2PA (Coalition for Content Provenance and Authenticity) manifests to embed immutable metadata into images, documenting their origin and the specific "ingredients" used in their creation.
By binding these three elements, EKILA prevents "copy-minting" and ensures that if an artist sells an NFT, the rights to future royalties for that image's use in AI training automatically transfer to the new owner.
Visual Attribution Pipeline
The core technical innovation of EKILA is its high-performance computer vision pipeline designed to prove which training images contributed to a generated output. Unlike standard semantic models like CLIP, which provide a "blunt" global similarity score, EKILA focuses on local, patch-level correlations to identify instances where an AI has memorized specific textures, objects, or artist signatures.
The pipeline consists of three stages:
- Patchified Fingerprinting: Images are decomposed into patches and processed through a ResNet-50 encoder trained via contrastive learning. This creates compact "fingerprints" that allow the system to detect granular details.
- Match Verification: A 3-layer Multi-Layer Perceptron (MLP) analyzes the correlation between query patches and retrieved training patches to filter out false positives.
- Credit Apportionment: A mathematical formula distributes a score across the top-K matching images, scaling royalty payments accordingly.
Experimental Results and Performance
The researchers evaluated EKILA against massive datasets, including LAION-400M and Adobe Stock's IPF-Stock. The system demonstrated significant improvements over existing baselines:
- Attribution Accuracy: On the LAION dataset, EKILA achieved a Recall@1 of 83.33%, vastly outperforming ViT-CLIP's 56.25%.
- Precision: In patch-level attribution, the system reached a mean Average Precision (mAP) of 72.73%, nearly 13 points higher than the CLIP baseline.
- Robustness: The fingerprints remained resilient even when generated images were subjected to degradation or modification.
Conclusion and Future Outlook
EKILA provides a technical blueprint for a formal model of consent and paid reuse in the GenAI era, analogous to the evolution of sampling rights in the music industry. While challenges remain regarding the computational scalability of indexing the entire internet and the need for supporting legal frameworks, EKILA demonstrates that automated, decentralized attribution is a viable path toward a more equitable digital creative economy.
Summary updated