Machine Learning Process Detection via Classifier Metrics
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing antivirus programs do not effectively detect and prevent malicious machine learning processes from stealing data or harming computing devices, particularly during training processes, as they do not specifically address machine learning training processes as a method to disrupt or access sensitive information.
Innovation Solution
A system and method that detect machine learning processes, whether malicious or not, by generating and analyzing process metrics on computing devices, using a classifier to determine the maliciousness of training code, and alerting users or stopping the process, which can operate with or without access to the training code, and includes distributed and local training approaches.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional antivirus programs are used to detect malicious processes, then general malware detection is provided, but machine learning training processes cannot be effectively detected or prevented from stealing data
Solution Approach 1:
The system performs preliminary actions by collecting process metrics and training data before actual machine learning training processes execute. It builds classification models in advance that can identify malicious training processes, enabling detection before data theft occurs rather than reacting after infection
Solution Approach 2:
The patent introduces process metrics as an intermediary layer between the machine learning training process and the detection system. Instead of directly analyzing training code or models, the system monitors intermediate metrics such as CPU usage, memory consumption, and system calls that characterize malicious training behavior, enabling indirect but effective detection
2Reliability
If machine learning processes are monitored and analyzed in real-time, then data theft prevention is improved, but system performance and resource consumption increase
Solution Approach 1:
The system applies partial monitoring by selectively collecting only specific process metrics that are most indicative of malicious training behavior, rather than monitoring all possible system parameters. This reduces the overhead of data collection while maintaining effective detection capability
Solution Approach 2:
The patent creates copies of process metrics and training data characteristics to build classification models, rather than directly interfering with or slowing down the actual machine learning training processes. The detection system operates on replicated data structures, leaving the original training workflows uninterrupted
3Measurement precision
If classification models are trained to distinguish malicious from non-malicious training code, then detection accuracy improves, but the complexity of the system increases
Solution Approach 1:
The detection system is segmented into distinct functional modules: process metric collection, data preprocessing, classification model training, and detection execution. Each module handles a specific aspect of the detection pipeline, making the overall complex system manageable through modular design and enabling independent optimization of each component
Data Source
AI summary
In some examples, the disclosure describes a device, comprising: a processor resource, and a non-transitory memory resource storing machine-readable instructions stored thereon that, when executed, cause the processor resource to generate a machine learning dataset, train and deploy a classifier using the machine learning dataset to determine whether a machine learning training process is running on a first computing device and whether the machine learning training process is malicious or is not malicious, and send an alert to the first computing device based on results of the trained classifier deployment on the first computing device.