On-Premises Generator

Integrates differentially private stochastic gradient descent (DP-SGD) directly into the model training loop. The generator clips per-sample gradients and injects calibrated Gaussian noise within the local trusted execution environment, ensuring that the final synthetic dataset carries a mathematically provable privacy guarantee (parameterized by epsilon). No raw gradients or intermediate statistics are ever transmitted outside the private subnet, eliminating the risk of gradient leakage attacks.

Unlike flat-file generators, an enterprise on-premises engine preserves referential integrity across complex relational databases. It models parent-child dependencies and foreign key constraints to generate entire synthetic database instances. Key capabilities include:

• Conditional tabular GANs (CTGAN) for mixed-type columns
• Automatic detection of primary key and foreign key relationships
• Preservation of cardinality and join patterns across tables
• Generation of consistent transactional time-series data

The generator includes an embedded evaluation framework that quantifies the utility of synthetic outputs before release. It computes Wasserstein distance between real and synthetic distributions for continuous columns, Jensen-Shannon divergence for categorical variables, and runs propensity score matching tests where a discriminator attempts to distinguish real from fake records. A fidelity score above 0.95 indicates the synthetic data is statistically indistinguishable for downstream machine learning tasks.

The synthesis model and its training pipeline execute exclusively within a trusted execution environment (TEE) backed by hardware roots of trust. This cryptographically verifiable isolation ensures that even privileged system administrators or hypervisor-level attackers cannot inspect the model parameters or the real data in use. The TEE attests to its integrity via remote attestation protocols before any synthesis job begins, providing auditable proof that the generator has not been tampered with.

Before any synthetic dataset is released, the generator automatically computes a re-identification risk score for each synthetic record. It measures the distance to the nearest real record in the training set and flags records that are too close to real individuals. The system enforces k-anonymity constraints and applies automated data masking to high-risk cells. A configurable risk threshold allows privacy officers to define acceptable disclosure levels per use case.

The generator continuously monitors the statistical properties of incoming real data streams to detect distributional shift. When the underlying data distribution changes—due to seasonality, new product launches, or market shifts—the system triggers an alert and can automatically schedule retraining of the generative model. This prevents the production of stale synthetic data that no longer accurately represents the current real-world population, ensuring ongoing utility for model training and testing.

The degree to which a synthetic dataset accurately reproduces the statistical properties, joint distributions, and complex inter-attribute relationships of the original real-world data.

• Propensity Score Matching: Measures how well a classifier can distinguish real from synthetic records; lower discriminability indicates higher fidelity.
• Wasserstein Distance: A metric measuring the minimum cost of transforming one probability distribution into another, used to evaluate distributional similarity.
• Utility vs. Privacy Trade-off: On-premises generators must balance high fidelity against the formal privacy budget (epsilon).

A database constraint enforced during multi-table synthesis ensuring that foreign key relationships between generated tables are valid and consistent, preventing orphaned synthetic records.

• Parent-Child Tables: The generator must model and preserve the relational links between tables.
• Sequential Synthesis: Often requires generating parent tables first to provide valid key ranges for child tables.
• On-Premises Necessity: Essential for synthesizing complex enterprise databases with normalized schemas without breaking application logic.