September 2020 Summaries

Creating differentially private, synthetic versions of datasets using Gretel.ai can help meet compliance requirements for sensitive data management, such as HIPAA, PCI, GDPR, and CCPA, while also enabling quicker project initiation without a data processing agreement. By utilizing Gretel.ai's tools locally, whether on a cloud or on-premises setup, users can generate high-quality synthetic models and datasets. The process involves setting up a suitable computing environment, generating an API key for access to Gretel's premium features, and installing necessary dependencies like TensorFlow and Pandas within a virtual Python environment. Training a model on a dataset, typically requiring at least 5,000 rows of data, allows for the creation of a synthetic dataset that maintains similar correlations and insights to the original data. Gretel.ai provides tools for validating synthetic data quality and offers a walkthrough guide on GitHub for a comprehensive understanding of the process. The use of synthetic data is anticipated to enhance machine learning models by reducing biases and improving generalization against unknown data.

The text explores the process of using machine learning and differential privacy to generate synthetic datasets from real-time e-bike ride-share data, specifically the General Bike-share Feed Specification (GBFS) from Los Angeles. It highlights the importance of anonymizing data to protect individual privacy, citing traditional methods like data anonymization and more advanced techniques like differential privacy. The goal is to create a synthetic dataset that retains the statistical utility of the original data while enhancing privacy, thus preventing re-identification attacks. The methodology involves training a generative neural network, using Recurrent Neural Networks (RNNs) and Generative Adversarial Networks (GANs), to produce realistic synthetic data. The text also discusses the challenges and benefits of applying differential privacy, which provides mathematical privacy guarantees crucial for compliance with regulations like GDPR and CCPA. The results show that the differentially private synthetic datasets can offer similar accuracy to the original data, with minimal memorization of individual data points, thus allowing secure data sharing.

Gretel has launched its Console Beta for creating data projects and uploading datasets to the cloud. Following user feedback, they have introduced a new feature called the project integration button. This allows users to access their project name, API key, and connection URI directly from any project page, making it easier to integrate with Gretel's SDKs or REST API. The company plans to continue improving its product based on user feedback.

Gretel Cloud has introduced a new entity stream view that allows users to explore record streams from tagged entities in their data projects. The platform uses advanced NER and NLP for automatic labeling of data, providing insights into the types of data being sent. With the entity stream feature, users can now browse records associated with an entity in a single view, enabling them to identify trends and sensitive fields. This functionality is accessible through Gretel Cloud Console, gretel-python-client SDK, and REST API. Users are encouraged to sign up or log in to experience the new feature and provide feedback for future updates.

Gretel Synthetics, in conjunction with Google Colaboratory's free GPUs, offers a streamlined approach to training machine learning models for generating fake, anonymized data with differential privacy guarantees. The latest version, 0.6.0, of the gretel_synthetics open-source synthetic data library introduces features such as Google SentencePiece support for unsupervised tokenization with adjustable vocabulary size and character coverage, and smart_open support for loading datasets from major cloud platforms like AWS, GCP, and Azure. Users can directly launch into Colaboratory to begin creating synthetic datasets. For further insights into anonymizing precise location data, the post references a detailed exploration of scooter ride-share data privacy issues and collaboration with Uber to address these concerns.

Gretel is a platform founded on the belief that data shouldn't be scary and aims to enable developers to safely share and collaborate with sensitive data. The company has built open-source SDKs, composable APIs, and an AI-based synthetic data library to address privacy concerns in working with sensitive data. Gretel's public beta is now available for free to any developer, offering guides on labeling and sharing datasets or generating synthetic datasets with differential privacy guarantees.

Gretel Outpost is a free integration architecture that automates risk assessment and exposure of data for security teams. Launching on August 25th, it enables developers to create safe data with cloud service and developer SDKs. The platform ingests alerts from common security tools, executes data collection surveys, and sends the data to Gretel for quick identification of PII and sensitive data exposure. Outpost consists of two components: Consumers that receive alerts from security tools and create survey request messages, and Surveys that collect data at a granular level and extract record-level information. The platform is currently in beta and can be accessed by contacting [email protected] for interested security or privacy organizations.

Gretel's REST APIs automatically build a metastore that makes it easy to understand what is inside of your data, making it accessible for developers familiar with the concept. The metastore helps in identifying primary identifiers and other potentially identifiable fields within the data. By using word embeddings and comparing field names, Gretel can infer relationships between schema field names without analyzing the field contents. In a case study involving GBFS free_bike_status.json feeds, it was found that some data producers were generating ephemeral bike_id values on a steady interval, while the jump_vehicle_name field had a one-to-one mapping with an actual vehicle. The privacy implications of this data are real, as ride reconstruction for specific individuals could be possible by observing bikes through the app and aggregated feed data. Gretel's disclosure on GitHub led to the removal of the data from the feeds within hours.

This post discusses the use of synthetic data to protect individual privacy while sharing electronic health records (EHR) among hospitals and healthcare organizations. Synthetic data is artificially manufactured rather than generated by real-world events, making it a promising technology for sharing knowledge in healthcare without compromising patient privacy. The article delves into training Gretel's open-source synthetic data library to generate EHR that protect individual privacy while capturing key statistical insights from the original source data. It provides an example dataset of de-identified emergency room discharge summaries and demonstrates how to use Gretel's ML-generated synthetic data to maintain both per-column distributions and field correlations in the training set, ensuring that individual health records are not memorized or replayed. The post concludes by expressing excitement about the potential of synthetic datasets for enabling safe data sharing with differential privacy guarantees in healthcare.