Privacy Enhancing Technology For Data: Protecting Live Data and AI Workloads

By: Rocco Alfonzetti On:

PETs (Privacy Enhancing Technologies) are increasingly vital in today’s digital landscape. PETs are essential for protecting individual rights, fostering trust, ensuring compliance, and supporting ethical practices in an increasingly interconnected world. Laws like GDPR, DORA and CCPA require companies to handle personal data responsibly. And PETs can help organizations comply with these regulations and avoid hefty fines. Besides being required by compliance, PETs also strengthen data protection and, in turn, bolster user trust.

Paperclip SAFE aligns with the key tenets of PET, allowing for data to be encrypted during ingestion and keeping it encrypted during use and throughout the entirety of the data lifecycle. There are a number of different PET techniques that are effective.

 

1. Data Anonymization and Pseudonymization

These techniques modify data to remove personally identifiable information (PII) while still allowing for analysis. Anonymization removes all identifiers, while pseudonymization replaces them with fictitious values.

Both Privacy solutions require robust data discovery and classification to identify sensitive data fields within your dataset.

Configuration uses techniques like redaction (complete removal), generalization (replacing specific values with broader categories), k-anonymity (ensuring each record is indistinguishable from at least k-1 others), and noise injection.

Most major DBMS solutions like Oracle, SQL Server, and MySQL offer built-in pseudonymization functionalities. These allow creating and managing pseudonymization keys for data transformation within the database environment.

Cloud providers like AWS, Azure, and GCP offer data anonymization and pseudonymization functionalities within their cloud storage solutions.

2. Data Masking

Data masking is a technique used to modify sensitive data within a dataset to protect privacy while still enabling data analysis and other purposes. It essentially creates a fake version of your real data that retains its structure and key characteristics but hides the actual values. AI data masking solutions are a critical tool for protecting sensitive information while enabling AI development and training.

Some of the available Data Masking solutions are:

  • IBM InfoSphere Data Masking
  • Informatica PowerCenter Data Masking
  • Precise DataFlex
  • Broadcom Test Data Manager
  • Denodo Data Masking (Open-source option)

 

3. Advanced Encryption for AI Privacy

 

3a. Homomorphic Encryption

This allows computations to be performed on encrypted data without decryption. While this ensures data privacy even while it’s being used by AI models, it can be computationally expensive for complex models.

  • OpenFHE:This is a popular open-source library led by a collaboration of industry and academic researchers. It offers various FHE schemes and focuses on usability and security. (https://github.com/openfheorg)
  • SEAL (homomorphic encryption library):Microsoft SEAL—powered by open-source homomorphic encryption technology—provides a set of encryption libraries that allow computations to be performed directly on encrypted data. This enables software engineers to build end-to-end encrypted data storage and computation services where the customer never needs to share their key with the service.(https://www.microsoft.com/en-us/research/project/microsoft-seal/)
  • IBM Cloud HEaaS (Homomorphic Encryption as a Service):This offering allows experimentation and development with FHE in a cloud environment. (Availability details are best obtained from IBM directly)
  • Enveil:This company offers a variety of homomorphic encryption solutions, including secure machine learning capabilities. https://www.enveil.com/

 

3b. Searchable Symmetric Encryption (SSE)

Imagine a vast medical archive with encrypted patient data. SSE creates a searchable index for this archive, allowing researchers to find specific medical images (data points) using keywords without decrypting everything.

Users can search for a particular disease using either the disease name or a secret code (search token), and the system locates relevant content while keeping the patient data itself encrypted. Researchers can then decrypt only the specific content they need with a master key. (Think of finding relevant medical content in a secure library without needing to open every file.)

  • Cloud storage providers with SSE capabilities:Some cloud storage providers, like Google Cloud Storage or Amazon S3, offer functionalities that incorporate simple SSE concepts. These services might not explicitly advertise “SSE” but provide features like client-side encryption with limited searchable indexes.
  • Paperclip SAFE®:is a data security platform (both SaaS and on premises) offering “always-encrypted” functionalities that interconnect with existing operational applications. incorporates SSE principles with patented shredding technology for searchable AES 256 encryption while data remains encrypted at rest and in use. (https://paperclip.com/safe/)
  • Pysearchable (Open Source Python library):This library provides various SSE implementations in Python. (https://github.com/ko1o/PYSearch)

 

4. Secure Multi-Party Computation (SMPC)

This enables multiple parties to collaboratively analyze data without revealing their individual datasets. This is useful for institutions working together on AI projects without compromising sensitive information.

o    OpenSMPC: (https://github.com/calcom/cal.com) – A popular open-source library offering various SMPC functionalities.

o    MP-SPDZ: (https://github.com/data61/MP-SPDZ) – Another well-regarded open-source library for secure multi-party computation.

o    TF Enclave: (https://github.com/tensorflow/tensorflow) – A library from Google focused on integrating SMPC functionalities with TensorFlow, a popular machine learning framework.

 

5. Federated Learning

In this approach, AI models are trained on local devices holding individual data. Only the model updates, not the raw data itself, are shared for central aggregation. This keeps user data private while allowing for collaborative training.

Federated learning is especially useful in the following categories.

  • Mobile Phone App Development
  • Healthcare and Medical Research
  • Internet-of-Things (IoT) and Smart Devices
  • Additional Privacy-Preserving AI Systems

 

6. Local Differential Privacy

A variant of differential privacy where noise is added to data on individual devices before it’s used for training, further enhancing privacy for federated learning applications. Privacy-Preserving Support Vector Machines (SVMs): These algorithms are specifically designed to train SVM models while protecting the privacy of the data used. SVMs are a type of machine learning model used for classification tasks.

  • Secure K-Nearest Neighbors (KNN): Similar to privacy-preserving SVMs, these techniques allow training KNN models (used for classification and regression tasks) without compromising data privacy.
  • Differential Privacy:This is a mathematical framework that ensures any information learned from a dataset cannot be linked back to a specific individual. It adds noise to the data in a controlled way, protecting privacy while still allowing for useful insights.
  • Research and Development Efforts
  • Several tech giants like Apple, Google, and Microsoft are actively researching and developing differential privacy techniques for their AI products.
    • Apple has mentioned using differential privacy in some of its features like keyboard usage statistics and health data collection on iPhones. However, specifics about the products themselves are not always publicly available.
  • TensorFlowPrivacy: (https://www.tensorflow.org/responsible_ai/privacy/guide)
    • This library from Google offers functionalities for incorporating differential privacy into TensorFlow, a popular machine learning framework.
  • OpenDP: (https://github.com/opendp/opendp) This is an open-source project providing various tools and libraries for implementing differential privacy in machine learning.
  • Transparent and Explainable AI (XAI): XAI techniques help users understand how AI models arrive at their decisions. This fosters trust and allows for identifying potential privacy risks within the AI system. Libraries like TensorFlow Explainable AI (TF-XAI) or DARPA’s Explainable AI Toolkit (XAI Toolkit) offer tools and techniques for building and explaining machine learning models.

 

The landscape of data protection today is a complicated web of technologies and half measures. Paperclip SAFE completes the circle of security that most other solutions ignore—protecting plaintext data in use. Paperclip SAFE is deployed across all Paperclip services, providing protections for customers who are required to provide due care to the data lifecycle.

Our mission with SAFE is to rid the world of plaintext data breaches and secure critical data across all verticals. Join us in our mission by learning more about Paperclip SAFE and encryption-in-use technology.