The fields of machine learning and data science have made tremendous strides, with the increasing use of artificial intelligence (AI) in various applications. However, as the amount of data generated and collected continues to grow, concerns over data privacy have become a critical issue especially in cases where data needs to be stored on the cloud. In response, researchers and industry professionals have developed techniques for privacy-preserving machine learning (PPML), a method of training machine learning models without exposing sensitive data.
PPML involves several techniques such as differential privacy, federated learning, homomorphic encryption, and secure multi-party computation. These techniques help to protect the privacy of sensitive data by ensuring that the data is processed and used in a secure and private manner, while still allowing models to be trained and optimized.
On the other hand, MLOps, which stands for Machine Learning Operations, is a set of best practices that aim to standardize, streamline, and automate the machine learning pipeline, from data ingestion to model deployment. It involves managing data, building and testing models, deploying and monitoring models, and managing infrastructure, among other tasks.
Combining the two areas, privacy-preserving machine learning (PPML) and MLOps, can offer significant benefits for organizations seeking to leverage AI while still preserving the privacy of their sensitive data. This approach involves building a secure, end-to-end machine learning pipeline that incorporates privacy-preserving techniques into every stage of the pipeline.
To face this issue we have developed HEflow: A Privacy-Preserving Machine Learning Lifecycle Platform
HEflow is a platform to streamline privacy-preserving machine learning development, including tracking experiments, packaging code into reproducible runs, and sharing and deploying encrypted models. Built on top of MLflow, Seldon MLServer and OpenMined TenSEAL, HEflow offers a set of lightweight homomorphic encryption APIs that can be used with any existing machine learning application or library (scikit-learn, Keras, TensorFlow, PyTorch, etc), wherever you currently run ML code (e.g. in notebooks, standalone applications, or the cloud).
Homomorphic Encryption (HE)
Homomorphic encryption differs from typical encryption methods in that it allows computation to be performed directly on encrypted data without requiring access to a secret key. The result of such a computation remains in encrypted form, and can at a later point be revealed by the owner of the secret key. This ground-breaking technology has enabled industry and government to provide never-before enabled capabilities for outsourced computation securely.
Homomorphic encryption workflows, for privacy-preserving machine learning, involve three entities:
- an ML model owner,
- a cloud server that performs model inference on HE encrypted data using the pre-computed ML model, and
- a user who sends confidential data to the cloud for model inference.
In all cases, the cloud should learn nothing about the underlying encrypted data.
Privacy-Preserving Machine Learning Operations (PPMLOps)
PPMLOps is a set of processes and automated steps to manage code, data, and encrypted models. This section describes a typical PPMLOps workflow.
PPMLOps using HEflow
- Experiments
Data scientists develop, train and tune the model on the production data, then they encrypt ① and register it with the Encrypted Model Registry. Model quality is evaluated by testing on held-out production data. This pipeline can be triggered by code changes or by automated retraining jobs. - Encrypted Model Registry
Autologging saves a record of the training and evaluation process, which includes model metrics, parameters, tags, and the encrypted model itself. When training and hyperparameter tuning are complete, the data scientist registers the final encrypted model artifact in the Encrypted Model Registry for the production environment. This records a link between the encrypted model and the code used to generate it. - Encrypted Model Serving
A continuous deployment (CD) process takes new encrypted models and deploys ② them for low-latency online serving (APIs). Options include cloud provider serving endpoints, or custom serving applications. - Invocations
The serving system loads the Production encrypted model version from the Encrypted Model Registry. For each request, it scores the encrypted data ③, and returns encrypted predictions ④.
What is HEflow?
HEflow is an open source platform developed by InAccel to help manage the complete privacy-preserving machine learning lifecycle with enterprise reliability, security and scale. It tackles four primary functions:
🛠️ Encrypted Model development
Accelerate and simplify privacy-preserving machine learning lifecycle management with a standardized framework for developing production-ready PPML models. With HEflow, you can bootstrap PPML projects, perform rapid iteration with ease and ship high-quality encrypted models to production at scale.
📋 Experiment tracking
Run experiments with any ML library, framework or language, and automatically keep track of parameters, metrics, code and encrypted models from each experiment. By using HEflow, you can securely share, manage and compare experiment results along with corresponding artifacts and code versions.
🧩 Encrypted Model management
Use one central place to discover and share PPML models, collaborate on moving them from experimentation to online testing and production, integrate with approval and governance workflows and CI/CD pipelines, and monitor PPML deployments and their performance. HEflow facilitates sharing of expertise and knowledge, and helps you stay in control.
📦 Encrypted Model deployment
Quickly deploy production encrypted models for batch inference or as gRPC homomorphic encryption APIs using built-in integration with Docker containers or KServe. With HEflow, you can operationalize and monitor production encrypted models to scale based on the business needs.
Installing
Install HEflow from PyPI via pip install heflow
Official HEflow Docker Image
The official HEflow Docker image is available on Docker Hub at https://hub.docker.com/r/inaccel/heflow.
# Pull the latest version
docker pull inaccel/heflowHEflow is open source and you can find more info here:
