The authors propose DMVM (Decentralized Multi-task Valuation via Model Merging), a framework that quantifies dataset contributions for multi-task models without retraining or sharing raw data. By leveraging task arithmetic to infer marginal utility from model combinations, the method addresses the computational costs and privacy constraints of traditional valuation approaches.
- Bypasses retraining and data sharing by inferring contributions directly from model parameter space.
- Provides a secure aggregation protocol for collaborative valuation without revealing individual model parameters.
- Includes theoretical error bounds characterizing approximation quality.
- Validated through experiments on computer vision and natural language processing tasks.
This approach offers a scalable and computationally efficient solution for fair data marketplaces in decentralized environments with strict privacy requirements.