Generating Privacy-Preserving Synthetic Data

Privacy-preserving data sharing is central to many Army needs: collaborating with partners in multi-domain operations, evaluating 3rd party vendors with realistic datasets, AI model building, and incident reproduction. Rockfish Data proposes to develop a new Generative Recording And Privacy-Preserving Replay (GRAPPR) system. The proposed GRAPPR framework uses cutting-edge research from Carnegie Mellon University on generative machine learning models to produce synthetic time series data for a broad range of applications. GRAPPR will be able to ingest different types of time-series datasets to create compressed snapshots, which can later be used to generate high-fidelity, privacy-preserving synthetic datasets for downstream use cases (e.g., sharing data with partners, testing vendor products in pilot deployments, developing more advanced AI model building capabilities). As a result, this system offers novel synthetic data generation capabilities for Army relevant use cases (e.g., tactical network traces, sensor logs, intrusion detection logs), which will be critical for future multi domain operations and which directly address the acquisitions program/platform requirements. During Phase I, Rockfish will demonstrate the feasibility of GRAPPR by developing a proof-of-concept end-to-end prototype for flexible data ingestion, compressive snapshotting, and privacy-preserving data generation In Phase II, Rockfish plans to add advanced capabilities for scalable private-cloud deployment to ingest larger datasets and support greater flexibility in user-defined fidelity metrics for synthetic data generation. Anticipated Benefits/Potential Commercial Applications of the Research or Development: Time series data is the cornerstone of any application - in the military and private sector alike. There is a lack of realistic data for model building, training, testing, prediction, simulation and incident response due to storage and/or privacy constraints. The Rockfish Data Generative AI model record-and-replay solution can be used to generate high-fidelity, privacy-compliant synthetic data. By generating periodic (e.g., daily) snapshots of available data, the Rockfish Data technology keeps a compact model representation of available data. These models and snapshots can then be used to create synthetic data for a particular period of time (e.g., a day, week, or month). This can then be used for machine learning model training, evaluation, and regression testing. Certain signals in the generated data can be amplified such as anomalies, threat vectors and fraud for stress testing. Certain fields can be modified in the generated data, while keeping other characteristics of the data, to enable simulations and what-if analysis. Moreover, the high-fidelity synthetic data allows for realistic workload for vendor evaluation and incident response.