Malware Detection Model Alignment via Ensemble Inversion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Anti-malware services on client devices face inconsistencies in malware detection due to different machine learning classifier models, leading to time-consuming updates and privacy concerns when transferring file samples, especially with the need to comply with privacy regulations.
Innovation Solution
A method involving the reception of probability scores from client devices, building a training dataset, and training a malware-detecting linear regression ensemble machine learning model specific to each identifier, which is then sent back to client devices for incorporation into malware-detecting services, reducing the need for file transfers and enhancing privacy by analyzing telemetry data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If file samples are transferred from client machines to backend locations for model updates, then machine learning models can be updated, but the process takes a long time and raises privacy concerns
Solution Approach 1:
Instead of transferring file samples from client to backend for model updates, the patent inverts the approach by having the backend send adjustment data (model updates) directly to clients. This eliminates the need for clients to upload large file samples, significantly reducing update time and addressing privacy concerns while maintaining model update capability.
Solution Approach 2:
The patent extracts only the essential adjustment data needed for model updates from the complex file sample data, and transfers only this compressed adjustment information to clients. This extraction approach maintains model update reliability while dramatically reducing data transfer time and privacy risks.
2Adaptability or versatility
If different machine learning classifier models are deployed, then diverse detection capabilities are achieved, but inconsistent conviction decisions occur for the same files
Solution Approach 1:
The patent merges multiple machine learning classifier models into an ensemble model that combines their detection capabilities. By integrating the strengths of different individual models into a unified ensemble approach, the system maintains diverse detection capabilities while achieving consistent conviction decisions across all clients.
Solution Approach 2:
The patent creates a universal adjustment data format that can be applied across different machine learning classifier models and client devices. This universal approach ensures that model updates maintain consistency and reliability across diverse detection capabilities, allowing the same adjustment data to improve multiple models uniformly.
3Reliability
If file samples are transferred to backend locations, then model training data is obtained, but privacy regulations are violated
Solution Approach 1:
Instead of transferring actual file samples that contain sensitive user information, the patent creates and transfers copies in the form of anonymized adjustment data and model updates. This copying approach maintains model training capability while eliminating privacy violations by removing all personally identifiable information from the transferred data.
Solution Approach 2:
The patent introduces adjustment data as an intermediary that mediates between the need for model training and privacy protection. This intermediary form of data contains the essential training information needed to improve models while acting as a barrier that prevents direct exposure of sensitive file samples, thereby complying with privacy regulations.
Data Source
AI summary
The disclosed computer-implemented method for producing adjustments to malware-detecting services may include (1) receiving, from a plurality of malware-detecting services executing on a plurality of client computing devices, a respective plurality of probability scores with corresponding model identifiers for an analyzed file and a plurality of respective identifiers describing the malware-detecting services, (2) building a training dataset from at least a portion of the received plurality of probability scores with corresponding model identifiers, and (3) performing a security action including (A) training, with the training dataset, a malware-detecting linear regression ensemble machine learning model that is specific to an identifier in the plurality of identifiers and (B) sending the trained linear regression ensemble machine learning model to one of the plurality of malware-detecting services executing on one of the client computing devices. Various other methods, systems, and computer-readable media are also disclosed.