<p>In the context of the accelerated marketization of data assets, existing privacy-preserving computation techniques face substantial limitations in ensuring the principle of “data availability without visibility.” Existing privacy-preserving techniques face significant limitations: differential privacy often degrades data utility due to the introduction of noise; homomorphic encryption imposes considerable computational overhead; and conventional generative adversarial networks (GANs) tend to compromise critical data characteristics. To address these challenges, this study introduces a novel privacy-preserving generative framework, GQ-ANGAN, which integrates graph-based Q-learning (GQ) with adaptive norm-constrained generative adversarial networks (ANGAN) in a multi-layered modular architecture specifically designed for numerical data. Specifically, the upper-layer GQ module is designed to protect discrete attributes by incorporating graph-structured representations into Q-learning, enabling dynamic optimization of privacy parameters and high-fidelity simulation of discrete features. The lower-layer ANGAN module targets continuous data and introduces adaptive normal distribution constraints to enhance the structural and statistical similarity between the generated and original data. A bidirectional parameter feedback mechanism connects the two modules, facilitating collaborative optimization and enabling flexible trade-offs between privacy strength and data utility across diverse application scenarios. Extensive experiments demonstrate that GQ-ANGAN generates synthetic data with strong privacy-preserving security and high consistency in statistical properties, while maintaining strong performance in both machine learning and deep learning tasks. Finally, the comprehensive evaluation with representative privacy algorithms highlights our framework’s capability to achieve an effective and scalable balance between robust privacy protection and practical data usability.</p>