Abstract
In the digital economy, data is traditionally considered a non-rival asset, meaning its consumption by one party does not diminish its value for others. This property creates a "data-sharing dilemma" where firms are hesitant to share strategic assets for fear of losing competitive advantages. This paper introduces the Encapsulation-Forging (EnFo) framework, a novel approach to data synthesis that transforms non-rival data into "rival" products. By engineering synthetic data to be highly effective for a specific authorized model while remaining useless to unauthorized ones, EnFo provides a mathematical "lock" on data utility. Experimental results across five datasets demonstrate that EnFo achieves high predictive accuracy with minimal data volume while offering superior privacy protection and resistance to unauthorized data augmentation.
Introduction
Data is a cornerstone of modern strategic competition. However, the non-rival nature of data—where it can be copied and repurposed without cost—often prevents mutually beneficial sharing between partners. Existing privacy-preserving methods, such as GANs or Diffusion Models, focus on symmetric utility, aiming to replicate the original data distribution as closely as possible. While effective for general-purpose synthesis, these methods fail to prevent unauthorized parties from extracting insights once they gain access to the synthetic data.
To address this, we propose a shift toward asymmetric utility. The goal is to create synthetic data that acts as a "rival" product: it is highly valuable to a designated "key" model but provides no utility to others. This approach leverages overfitting—traditionally viewed as a flaw—as a security feature to create a technical barrier against data repurposing.
The Encapsulation-Forging (EnFo) Framework
The EnFo framework operates through a two-stage process designed to "lock" the predictive power of a dataset into a specific model architecture.
1. Knowledge Encapsulation (Data to Model)
In the first stage, the data owner selects a target "key" model $\mathcal{A}_\theta$ (e.g., a specific regression algorithm) and trains it on the original dataset $D$. This process encapsulates the predictive relationships within the raw data into the model’s parameters $\hat{\theta}_D$.
2. Asymmetric Utility Forging (Model to Data)
The second stage involves a novel meta-learning optimization. Instead of training a model on data, the framework optimizes the synthetic data $\widehat{D}$ itself. The objective is to minimize the loss of the key model on the original validation set when trained on the synthetic data: $$\widehat{D} = \text{argmin}_{\widehat{D}} \frac{1}{N} \sum_{i=1}^N \mathcal{L}(\hat{f}(x_i), y_i)$$ By keeping the synthetic dataset extremely small and avoiding standard regularization, the data is "forged" to overfit the specific architecture of the key model. This ensures that the synthetic data fails to generalize to unauthorized models or external data augmentation.
Extensions for Flexibility
The framework includes two key extensions to balance security with practical utility:
- Statistical Alignment: Uses Wasserstein distance to ensure the synthetic data remains "mixable" for authorized partners who need to perform data augmentation.
- Model Adaptability: Introduces a term to reduce synthetic data complexity, allowing it to function across a pre-approved "toolkit" of models rather than a single architecture.
Experimental Validation
The EnFo framework was evaluated using five diverse datasets, including Monte Carlo simulations (CLV prediction, Friedman #3) and real-world marketing data (Instacart, Retailrocket, CDNOW).
Key Results
- Rivalry and Specificity: EnFo effectively "breaks" unauthorized models. On the Friedman dataset, switching from the designated model to an unauthorized one resulted in an 80% performance degradation.
- Sample Efficiency: The framework is remarkably efficient; on the CLV dataset, EnFo matched the accuracy of the full original dataset using only 1% of the data volume.
- Privacy and Resistance: EnFo outperformed traditional GANs and Diffusion models in privacy metrics, achieving Monte Carlo Attack Accuracy (MCAA) scores below 0.57. Furthermore, unauthorized attempts to "fix" the data via external augmentation actually increased prediction errors by up to 32%.
Conclusion and Future Work
The Encapsulation-Forging framework represents a paradigm shift in data sharing, moving security from physical or legal restrictions into the mathematical properties of the data itself. By creating "rival" synthetic data, EnFo enables secure collaboration without sacrificing competitive advantage.
Future research will focus on extending this framework to unstructured data types like text and images, as well as developing protocols for multi-party sharing environments where multiple owners contribute to a shared rival data pool. While the bi-level optimization process is computationally intensive, the resulting data minimization and security benefits provide a compelling solution for the modern data economy.
Summary updated