Leveraging arithmetic circuits for differentially private and cost-efficient synthetic data generation
- Publication Type:
- Thesis
- Issue Date:
- 2026
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
This thesis presents a synthetic data generation approach using Probabilistic Graphical Models (PGMs) as the underlying data generating process. Differential privacy is applied to protect the underlying reference data. To address the computational challenges of applying Differential Privacy to large cross-tables, this thesis proposes a novel decomposition scheme for Laplacian noise perturbation. Evaluation demonstrates that the proposed decomposition method reduces the time required for noise injection by a factor of three while maintaining the same accuracy as the standard method. After creating a generative model, synthetic data may be generated by sampling the model. To address the high computational cost of independent sampling for synthetic data generation, this thesis proposes two novel cost-efficient independent sampling methods to draw samples from a PGM, where the PGM is compiled to an arithmetic circuit. Empirical evaluation confirms these proposed samplers achieve the highest computational efficiency among other independent samplers. The final contribution of this thesis is the construction of a novel cost-efficient and differentially private synthetic data generator that maintains referential integrity. The proposed synthetic data generator had the best data generation time when evaluated against similar state-of-the-art generators using multi-table relational datasets as workloads.
Please use this identifier to cite or link to this item: