Time-Based Score Oscillation for ML Model Obfuscation

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Machine learning model outputs can be intercepted and manipulated by malicious actors through iterative modifications, leading to potential security breaches in applications like malware detection, where the model's classification can be manipulated to classify malicious files as safe.

Innovation Solution

Implementing a time-based oscillation function to obfuscate the output scores of classification models, using a combination of periodic functions with attenuation to create bounded noise, making it difficult for adversaries to reverse-engineer or predict the true score, thereby preventing iterative improvements in malicious file classification.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of operation

If machine learning model outputs are made publicly accessible for consumption, then application functionality is enabled, but security vulnerability increases allowing malicious actors to intercept and manipulate classifications

Engineering Contradiction:

Improvemodel output accessibilityVSAvoidsecurity vulnerability

Core Design Contradiction:

Ease of operationVSObject-affected harmful factors

Solution Approach 1:

The patent introduces a time-based oscillation function as an intermediary layer between the classification model and the consuming application. This mediator adds bounded noise to the true score, creating an obfuscated output that prevents direct manipulation while preserving the model's protective function. The oscillation function acts as a buffer that masks the true classification confidence without blocking the security benefit.

Inventive Principle:

Principle #24Intermediary (Mediator)

2Measurement precision

If model output scores are provided with high precision, then classification accuracy is improved, but susceptibility to iterative manipulation increases

Engineering Contradiction:

Improveclassification score precisionVSAvoiditerative manipulation vulnerability

Core Design Contradiction:

Measurement precisionVSObject-affected harmful factors

Solution Approach 1:

The patent transforms the output parameter from a precise static score to a dynamically obfuscated score with added temporal noise. By changing the parameter representation through the time-based oscillation function, the system maintains the informational value of the classification while removing the exploitable precision that malicious actors could use for iterative attacks.

Inventive Principle:

Principle #35Parameter changes

3Reliability

If score obfuscation is applied to prevent manipulation, then security is enhanced, but information fidelity for consuming applications may be reduced

Engineering Contradiction:

Improvesecurity against manipulationVSAvoidscore information fidelity

Core Design Contradiction:

ReliabilityVSLoss of information

Solution Approach 1:

The patent applies partial obfuscation by adding bounded noise rather than completely masking the score. The attenuation parameter controls the noise magnitude to ensure it provides sufficient security protection while maintaining enough information fidelity for the consuming application to make accurate decisions. This partial action balances security enhancement with information preservation.

Inventive Principle:

Principle #16Partial or excessive action

Data Source

PatentUS11580442B2Machine learning model score obfuscation using time-based score oscillations

Publication Date: 2023.02.14 CYLANCE INC

AI summary

An artefact is received. Features are later extracted from the artefact and are used to populate a vector. The vector is input into a classification model to generate a score. This score is then modified using a time-based oscillation function and is provided to a consuming application or process. Related apparatus, systems, techniques and articles are also described.