Geometry-aware synthesis
Pairwise structure for tabular diffusion.
Privacy-Preserving Machine Learning Research Scientist
David Turtora Zagardo
I build differentially private synthetic data systems, privacy attacks and evaluations, and deployable ML infrastructure for organizations that need useful models without exposing sensitive records.
ICML 2026
Sole-author, self-funded paper on geometry-aware tabular synthetic diffusion.
5+ years
Building privacy-preserving ML and privacy engineering systems.
$170K+
Privacy engineering consulting revenue in 2025.
7 cases
Expert witness analysis of tracking, pixels, cookies, and fingerprinting.
Private data generation
Accounting, clipping, noise, and release controls.
Attack evaluation
Membership and attribute inference evidence.
Fitness-vector privacy
Centered population signals with RDP accounting.
Selected work at the research-to-production boundary.
These projects show how privacy guarantees, model utility, attack evaluation, and deployment constraints connect in real systems.
Synthetic Data Research
Geometry-Aware Tabular Diffusion
Sole-author, self-funded ICML 2026 paper introducing pairwise geometric features for tabular diffusion, improving fidelity and utility with fewer parameters than transformer baselines.
Production DP
Private Synthetic Data Pipelines
Built tabular, time-series, and text synthesis systems with DP-SGD, RDP accounting, group privacy, private preprocessing, and on-prem deployment controls for sensitive enterprise workflows.
Privacy Risk Evaluation
Membership and Attribute Inference Attacks
Built attack tooling for language and image models, plus FAISS-based attribute inference dashboards for synthetic data risk, translating privacy failures into measurable evidence.
Algorithmic Privacy Research
DP-EGGROLL
Differentially private evolution strategies via centered fitness vector privatization. DP-EGGROLL combines clipping, Poisson subsampling, and RDP accounting, with practical classification non-inferiority to fast DP-AdamW on 22/25 AUROC endpoints.
Privacy engineering toolkit.
The throughline is practical privacy-preserving ML: mechanisms, accounting, attack evaluation, and governance-aware deployment.
Differential Privacy
DP-SGD, Laplace, Gaussian, Exponential, Sparse Vector, subsampling, shuffling, RDP, and zCDP.
Private ML Systems
PyTorch, Opacus, HuggingFace, PEFT, LoRA, vLLM, FAISS, diffusion models, and deployment packaging.
Risk Evaluation
Membership inference, attribute inference, reconstruction risk, utility testing, and synthetic data dashboards.
Privacy Governance
Privacy by Design, PIAs, OneTrust, cookie consent, tracking technology review, and expert reporting.
Publications and service.
Research outputs, open-source work, and service that reinforce the privacy and ML positioning.
2026
Geometry-Aware Tabular Diffusion
ICML 2026. Sole author and self funded.
Research
DP-EGGROLL: Differentially Private Evolution Strategies via Fitness Vector Privatization
Research note and code with experiments against fast DP-AdamW, tuned uncentered EGGROLL, and scalar DP-ZO.
2024
Blockwise Gradient Aggregation for Deep Learning
IEEE Digital Privacy. Algorithmic work on blockwise gradient aggregation and model privacy evaluation.
2024
A More Practical Approach to Machine Unlearning
arXiv paper cited 6 times.
Service
ACM CODASPY Program Committee
Program committee member for 2025 and 2026.
Experience.
A concise narrative of the roles behind the research and systems work.
Privacy Engineer, ML Research
Secludy AI
Built DP synthetic data systems, inference attacks, privacy accounting, on-prem deployment workflows, and private LLM fine-tuning infrastructure.
Owner
Green Willow Studios
Enterprise privacy engineering consulting across differential privacy, synthetic data, compliance infrastructure, open-source tools, and expert reporting.
Technical Privacy Consultant
WebXRay
Audited web tracking, pixels, fingerprinting, cross-site tracking, and cookie consent practices across large-scale web datasets.
Research Assistant
Carnegie Mellon University
Studied privacy preferences and health data granularity, with survey design, statistical modeling, and cross-cultural privacy analysis.
Build useful ML without exposing sensitive data.
I am available for privacy-preserving ML research, synthetic data systems, differential privacy consulting, and technical privacy analysis.
GitHub
Focus
Differential privacy, synthetic data, privacy attacks, LLM privacy, and privacy engineering.