ML Model Abstraction Layer for Runtime Efficiency
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models in network security face challenges in achieving runtime efficiency due to the inclusion of irrelevant training information, such as hyperparameters and metadata, which can lead to increased latency and processing overhead, particularly in cloud-based systems.
Innovation Solution
A machine learning model abstraction layer is introduced to decouple the training process from the runtime process by removing irrelevant information like hyperparameters and metadata, resulting in a lean, purpose-built model for runtime use, supporting any architecture and algorithm.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If training information (hyperparameters, metadata) is retained in the machine learning model, then the model can be retrained and adjusted, but runtime latency and processing overhead increase
Solution Approach 1:
The patent segments the machine learning model into two distinct parts: a compressed runtime representation stripped of training information for fast inference, and separate training artifacts stored externally for potential retraining. This segmentation allows the runtime model to operate with minimal overhead while preserving adaptability through external training data access.
Solution Approach 2:
The patent extracts and removes training information (hyperparameters, metadata, training datasets) from the deployed machine learning model. By taking out only the essential inference capabilities and discarding training-specific elements, the model achieves faster runtime performance while retraining capability is maintained through external access to training artifacts.
2Productivity
If training information is removed from the machine learning model, then runtime efficiency improves, but the model cannot be easily retrained or adjusted
Solution Approach 1:
The patent introduces an intermediary layer that separates the compressed runtime model from the training artifacts. This intermediary allows the stripped model to operate efficiently at runtime while maintaining the ability to retrieve original training information when retraining is needed, thus bridging the gap between efficiency and adaptability.
3Loss of energy
If a lean, purpose-built model is used for runtime, then processing overhead is reduced, but the model becomes architecture-specific and less versatile
Solution Approach 1:
The patent creates a universal compressed model representation that can be deployed across different architectures and algorithms. By removing architecture-specific training information while preserving the core inference logic, the model achieves both reduced processing overhead and maintained versatility across multiple platforms and model types.
Data Source
AI summary
Systems and methods include receiving a trained machine learning model that has been processed with training information removed therefrom, wherein the training information is utilized in training of the trained machine learning model; monitoring traffic, inline at the node, including processing the traffic with the trained machine learning model; obtaining a verdict on the traffic based on the trained machine learning model; and performing an action on the traffic based on the verdict.