Cloud ML Notebook Data Loss Prevention via Command Risk Scoring
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current data loss prevention (DLP) systems do not effectively monitor and prevent indirect data losses during feature engineering and model training processes in machine learning (ML) notebooks, where sensitive information can be transformed and represented as data proxies, leading to security risks similar to direct data losses.
Innovation Solution
A cloud-based DLP system using machine learning notebooks that monitors and classifies commands within ML notebooks, generates risk scores, and alerts when thresholds are exceeded, incorporating a rule engine for organization-level security controls and cell obfuscation to restrict access to sensitive data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional DLP systems are used to monitor data loss, then direct data losses can be prevented, but indirect data losses during feature engineering and model training cannot be detected
Solution Approach 1:
The system segments the monitoring process into distinct components: a plugin integrated within the ML notebook environment that captures cell executions, a command parser that analyzes individual commands, and a risk scoring system that evaluates potential data losses. This segmentation enables detection of indirect data losses in feature engineering that traditional monolithic DLP systems miss.
Solution Approach 2:
The system introduces a command parser as an intermediary between the ML notebook environment and the risk scoring mechanism. This parser translates complex ML commands into analyzable components, enabling the system to detect indirect data losses through transformations and visualizations that would otherwise be invisible to traditional DLP systems.
2Reliability
If monitoring of all commands in ML notebooks is implemented, then indirect data losses can be detected, but system complexity increases
Solution Approach 1:
The system applies partial monitoring by focusing only on commands that pose potential data loss risks. The command parser identifies and analyzes only relevant commands (e.g., transformation, visualization, export operations) rather than processing every command in the notebook, reducing system complexity while maintaining effective detection of indirect data losses.
Solution Approach 2:
The monitoring system applies different analysis depths to different commands based on their potential risk. High-risk commands like data export or transformation receive detailed analysis, while low-risk commands receive minimal or no analysis. This local quality approach reduces overall system complexity while maintaining high detection capability for critical operations.
3Reliability
If risk scoring and alerting is implemented for all notebooks, then data security is enhanced, but false positives may increase
Solution Approach 1:
The system dynamically adjusts risk scores based on multiple parameters including the type of command executed, the sensitivity of the data involved, the context of the notebook, and the user's permissions. This multi-parameter approach allows the system to distinguish between legitimate analytical operations and actual data loss risks, reducing false positives while maintaining high alert accuracy.
Solution Approach 2:
The system implements feedback mechanisms where alert outcomes and user responses are used to refine risk scoring. When users confirm or dismiss alerts, this information feeds back into the risk scoring algorithm, improving its ability to distinguish true risks from false positives over time, thereby enhancing alert accuracy without compromising data security.
Data Source
AI summary
Embodiments secure data on a cloud based network that comprises one or more machine learning (“ML”) notebooks. Embodiments monitor activity on each of the ML notebooks, the activity including one or more commands. Embodiments classify each of the commands, the classifying including generating input parameters. Based on the input parameters, embodiments determine a risk score for each of the ML notebooks. When the risk score exceeds a predetermined threshold, embodiments generate an alert.