Technical Summary: AME – A Multi-Type Contributor Attribution Framework in Generative AI Markets

1. Problem Statement

Generative AI ecosystems involve multi-stage, multi-party collaboration where value is created through heterogeneous contributions, including training data, base models, fine-tuning behaviors (e.g., LoRA), and prompts. Existing data valuation and revenue allocation mechanisms are insufficient for this environment because they are designed for homogeneous participants and single-stage production settings.

The paper identifies three core challenges in allocating value within these multi-stage pipelines:

1. Inaccurate Attribution: Current methods struggle to fairly quantify and aggregate contributions from diverse types of contributors (data, models, effort, prompts) across multiple stages.
2. Mismatched Mapping: Quantified contribution values cannot be directly linked to rights holders due to complex authorization structures, the separation of ownership and usage rights, and diverse licensing terms.
3. Insufficient Trust: Ensuring trustworthy execution of allocation algorithms is computationally expensive. While blockchain offers transparency, the cost of verifying complex generative AI workloads (training, inference, attribution) via full verification (e.g., Zero-Knowledge Proofs) is often prohibitive.

2. Methodology: The AME Framework

The authors propose AME (Attribution–Mapping–Execution), a unified framework consisting of three logical layers designed to address the challenges sequentially.

Layer 1: Attribution Layer (MMShapley)

This layer introduces MMShapley, a multi-type, multi-stage attribution algorithm based on the Shapley value.

• Multi-Stage Cooperative Game Modeling: The pipeline is modeled as a directed acyclic graph ($G$) with layers ($L$). Contributors are partitioned into four sets: training data ($N_{data}$), models ($N_{model}$), builder effort ($N_{effort}$), and prompts ($N_{prompt}$).
• Multi-Type Utility Function: To evaluate semantic contributions, the framework uses a CLIP-based similarity function between generated images and a target image. The utility of a coalition is defined as the expected similarity.
• Handling Incompleteness: The utility function accounts for coalition completeness. For instance, a coalition with only data but no model yields zero utility, while a coalition with data and a default model yields partial utility.
• Small-Coalition Penalty: To prevent "free-riding" by weak contributors in small coalitions, a penalty operator ($\delta_{size}$) is applied. If a contributor's independent value is below a threshold ($\theta$), their weight in small coalitions is significantly reduced.
• Inter-Stage Propagation: A global retrospection operator propagates contribution values from downstream to upstream layers. A distance decay function ($\delta_{dist}$), derived from information theory (considering channel noise and information bottlenecks), reduces the weight of contributions as their distance from the final output increases.
• Complexity Optimization: To avoid the exponential complexity ($O(2^n)$) of standard Shapley calculations, the framework leverages the additivity of LoRA modules and the hierarchical structure of pipelines, decomposing the problem into independent intra-layer subgames.

Layer 2: Mapping Layer (Rights Credentialing)

This layer translates abstract contribution values into executable economic rights using a decentralized, NFT-based mechanism.

• Decoupling Ownership and Usage: Each contribution source is minted as an ERC-721 NFT. The framework separates Ownership NFTs (full control) from Usage-Right NFTs (authorized permissions).
• Programmable Licensing: Usage-right NFTs contain metadata for dynamic terms, including authorization types (fixed fee, revenue share), validity periods, and profit-sharing ratios.
• Revenue Allocation Formula: The final allocation for a rights holder $j$ is calculated by summing the revenue shares derived from their authorized contribution sources (weighted by MMShapley values) and any fixed fees. This supports nested authorization and multi-level profit sharing.

Layer 3: Execution Layer (Trustworthy Protocol)

This layer ensures the implementation of the allocation scheme through a dual-layer protocol combining smart contracts and a behavior verification mechanism.

• Smart Contract Workflow: Contracts manage the entire lifecycle: verifying ownership/usage rights, scheduling compute nodes, training/inference execution, and settling fees.
• Staking, Auditing, and Reputation: To ensure trustworthiness without the high cost of full verification, the system employs:
  • Staking: Nodes must lock a stake ($S$) to participate.
  • Random Auditing: Results are spot-checked with probability $p$.
  • Slashing: Failed audits result in a portion of the stake ($\gamma S$) being slashed.
  • Reputation: A reputation score ($\rho_t$) is updated via an exponential moving average based on audit outcomes. High reputation correlates with higher future revenue opportunities.
• Incentive Compatibility: Theoretical analysis demonstrates that the combination of immediate economic penalties (slashing) and long-term reputation loss creates an incentive-compatible environment where rational nodes prefer honest behavior, even under low audit frequencies.

3. Key Contributions

1. MMShapley Algorithm: A novel attribution method that extends Shapley values to heterogeneous, multi-stage generative AI pipelines, incorporating coalition-size penalties and inter-layer distance decay.
2. Unified AME Framework: The first framework to integrate contribution attribution, rights mapping (via decoupled NFTs), and trustworthy execution into a single workflow for generative AI markets.
3. Cost-Efficient Execution Protocol: A mechanism that achieves trustworthy allocation by combining smart contracts with a staking-auditing-reputation system, avoiding the prohibitive costs of full cryptographic verification for AI workloads.

4. Experimental Results

The paper validates the framework through image generation experiments using Stable Diffusion 1.5, LoRA fine-tuning, and human evaluation.

• Attribution Reasonableness:

  • Human Alignment: AME's attribution rankings showed perfect consistency (Spearman $\rho=1.0$, Kendall $\tau=1.0$) with human baselines in intra-stage evaluations, significantly outperforming Uniform Allocation and the Daraxide framework.
  • Global Attribution: In global evaluations, AME achieved the lowest Normalized Mean Absolute Error (NMAE = 0.104) and highest correlation with human intuition compared to baselines.
  • Style Absence: In scenarios where a specific style (e.g., Claymation) was absent, AME successfully suppressed the contribution of the irrelevant LoRA model to near-zero (0.012), whereas other methods assigned it significant weight.
  • Ablation Studies: Experiments confirmed that both distance decay and coalition-size decay operators are critical for minimizing NMAE and improving ranking consistency.

• Execution Feasibility:

  • Gas Consumption: Deployed on an Ethereum testnet, the framework's high-frequency operations (e.g., authorization, task creation) consumed between 51,000 and 210,000 Gas, which is deemed acceptable for supporting active collaborative ecosystems.
  • Throughput: The system achieved theoretical throughputs of up to 41.5 TPS for authorization settings, sufficient for current platform needs.

• Incentive Compatibility:

  • Simulation: Monte Carlo simulations demonstrated that the combination of staking, random auditing, and reputation mechanisms effectively drives the population honesty rate to ~95.7%, significantly outperforming systems with only staking or only auditing.
  • Safety Margins: The analysis identified a "probabilistic safety corridor," showing that a stake-to-revenue ratio ($S/R$) of 7.0 with a 10% audit probability achieves >99.9% safety, verifying the robustness of the incentive design.

5. Significance and Claims

The paper claims to provide an initial foundation for value assessment and revenue allocation in generative AI data markets. It addresses the fragmentation of existing research by offering a unified solution that connects:

• Valuation: Fairly quantifying heterogeneous contributions.
• Rights: Dynamically mapping values to rights holders via programmable licenses.
• Execution: Ensuring trustworthy and cost-effective settlement.

The authors emphasize that while the current evaluation focuses on image-generation pipelines and relies on simulations for incentive validation, the AME framework establishes a necessary theoretical and practical basis for future research into fair and sustainable generative AI economies. They acknowledge limitations, including the focus on image generation, the need to address jurisdiction-specific legal requirements, and the lack of real-world deployment validation.