Machine Learning (ML) model pipeline with obfuscation to protect sensitive data therein
Provided is a system and platform for Machine Learning (ML) based Data Discovery and Classification. The system and platform comprising components of a user console, a ML agent, and a ML data engine. By way of a ML pipeline, sensitive data is obfuscated that would otherwise by in the clear when transmitted to a centralized server. The ML model pipeline decouples embedding from model training. In a first step, the ML Agent runs on data endpoint machine or proxy to convert clear text data to embedding vectors. In a second step, the ML data engine runs on a centralized server to train models using the embedding vectors. The separation of pipeline components and respective handling of workflow requests and messages associated therewith prevents the transfer of clear data in the open. Other embodiments disclosed.
1 . A system for Machine Learning (ML) based Data Discovery and Classification (DDC), the system comprising components of:
a user console ( 100 ), running on an endpoint machine at a branch location managing and accessing data in a security zone under a security policy, said data in part is private or sensitive, for
processing user requests for data discovery and classification (DDC) on the endpoint machine;
ingesting user requests into a ML pipeline for embedding, training and deploying ML models on said data produced via DDC;
displaying classified data category and identified sensitive entities of said data on the endpoint machine by way of the ML pipeline;
a ML agent ( 200 ), communicatively coupled to the user console, also running on the endpoint machine and residing at the branch location, for
polling said user requests by way of the ML pipeline;
scanning the endpoint machine for said data responsive to user requests,
embedding said data to produce an embedding vector that is ingested into the ML pipeline instead of clear data;
applying ML models to the data scanned;
a ML data engine ( 300 ), communicatively coupled to the user console and the ML agent, and not residing at the branch location with said security zone, for
handling user requests from user console and ML agent on the ML pipeline; receiving said embedding vector on the ML pipeline;
labeling said embedding vector with labels responsive to user annotations; persisting, training, updating, and applying ML models for use by the ML agent,
wherein said components execute on a computational device comprising one or more processors and memory coupled to the one or more processors, wherein the memory includes computer instructions which when executed by the one or more processors causes the one or more processors to perform said operations.
2 . The system of claim 1 , wherein the embedding extracts parameters from the security zone and the security policy as features and adapts accordingly to comply with user-privacy laws and protect against leakage of sensitive data.
3 . The method of claim 2 , wherein the embedding provides a solution for compliance with user-privacy laws and protects against leakage of sensitive data.
4 . The system of claim 2 , wherein the ML pipeline codifies and automates a ML workflow for data discovery and classification in the user console, the ML agent and the ML data engine, and said workflow provides for processing requests, handling messages, data scanning, embedding data, sampling, annotating, cleaning, training models, persisting models, and applying models.
5 . The system of claim 2 , wherein one or more parameters of the security policy that include purpose, audience, objectives, authority, data classification, data sensitivity, and encryption policy are extracted as a first set of features of the embedding vector along with one or more parameters of the security zone that include type, level, location, name, trust, certificates, and data traffic rules extracted as a second set of features.
6 . The system of claim 1 , wherein sensitive entities includes private information or private data including, but not limited to: URL, entity, gender, sex, race, ethnicity, occupation, marital status, citizenship, or phone numbers.
7 . The system of claim 1 , wherein said ML model is one of:
an ML embedding model to generate an embedding vector from said data;
a ML training model based on annotation labels and embedding vectors; and
a ML prediction model to classify said data and identify sensitive entities within said data.
8 . The system of claim 7 , wherein the ML pipeline supports workflows consisting of:
an ML model embedding scan workflow;
an ML model on-line training workflow; and
an ML model prediction scan workflow.
9 . The system of claim 8 , wherein ML model package catalog management is provided via:
the user console ( 100 ), that by way of the ML data engine ( 300 ),
processes user input for uploading, downloading, deleting and viewing ML model packages for ML embedding models, ML training models and ML prediction models; and
displays package information and utilities on the ML pipeline for performing each of the three workflows.
10 . The system of claim 8 , wherein for the embedding scan workflow:
the user console ( 100 ) responsive to user input submits an embedding scan request on the ML pipeline;
the ML agent ( 200 ) polls and retrieves the embedding scan request on the ML pipeline and downloads downloads, instantiates, and runs the ML embedding model;
the ML agent ( 200 ) scans a folder identified in the user request, the folder containing data on the endpoint machine; and
the ML agent ( 200 ) embeds the data using the ML embedding model and submits the embedding vector, with obfuscated data therein, to the ML pipeline.
11 . The system of claim 10 , wherein for the ML model on-line training workflow: the user console ( 100 ), responsive to the embedding scan workflow, thereafter receives data labels submitted by the user for the embedding vector, and submits a model training request with the data labels on the ML pipeline; the ML agent ( 200 ) polls and retrieves the on-line training request on the ML pipeline with the data labels, and downloads, instantiates, and runs the ML training model; the ML data engine ( 300 ), samples multi-dimensional vector data points of the embedding vector into a number of groups by way of K-Means clustering; annotates the embedding vector in the groups with the user-provided data labels for classification for named entity recognition; applies a semi-supervised machine learning by way of K-Nearest Neighbors to propagate document labels to surrounding unlabeled data points; and applies supervised machine learning by way of a Neural Network to train the ML prediction model using embedding vector with labels as inputs.
12 . The system of claim 8 , wherein for the ML model prediction scan workflow:
the user console ( 100 ) handles a user request responsive to the user submitting a prediction scan request on the ML pipeline;
the ML agent ( 200 ) polls and retrieves the prediction scan request on the ML pipeline and downloads, instantiates, and runs a package for the ML model;
the ML agent ( 200 ) scans a folder for data identified in the user request and embeds the data using the ML embedding model from the package and submits the embedding vector to the ML pipeline;
the ML data engine ( 300 ) responsive to an interface call for prediction, retrieves the embedding vector from the ML pipeline, applies the ML prediction model from the package to each scanned file in the folder, and classifies the data into data categories and identifies sensitive entity in the data categories.
13 . The system of claim 8 , wherein communication on the ML pipeline for each workflow, by and between the user console ( 100 ), the ML agent ( 200 ) and the ML data engine ( 300 ), is via RESTful Application Programming Interfaces (APIs).
14 . The system of claim 7 , wherein said ML model consists of:
a metadata record consisting of ML model information, such as, identifier (ID), name, model type, national language, target file type, serialization file, executable binary, executable artifact and model description;
a serialization file that converts data objects of the ML model into a transmittable format over the ML pipeline; and
an executable binary that implements a predefined public call interface, which is invoked over the ML pipeline to interact with training, embedding and trained prediction models.
15 . The system of claim 1 , wherein said ML data engine ( 300 ) comprises:
a Deep Neural Network (DNN) to classify unstructured data; and
a Named Entity Recognition (NER) to detect sensitive entities within said unstructured data.