Machine Learning Input Variants for Privacy-Preserving External Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The risk of data leakage and loss of intellectual property is a concern when using externally-deployed foundation models like ChatGPT or ViT, as companies prefer to retain control over their proprietary data and avoid training on external servers.
Innovation Solution
A method involving pre-processing and post-processing of data on local devices to generate multiple variants of input and output data, respectively, using stochasticity of machine learning models to ensure privacy and control over data usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If externally-deployed foundation models are used for data processing, then computational resources and training effort are saved, but data privacy and intellectual property protection deteriorate
Solution Approach 1:
The patent segments the data processing workflow into multiple independent components: local pre-processing of input data to generate variants, external model inference on each variant, and local post-processing to aggregate results. This segmentation allows the organization to keep sensitive data locally while utilizing external computational resources, thus resolving the contradiction between saving computational resources and protecting data privacy.
Solution Approach 2:
The patent introduces processed data variants as intermediaries between the local system and the external foundation model. Instead of sending raw sensitive data directly to the external model, the system transforms it into multiple processed variants locally first. These variants serve as intermediaries that preserve the necessary information for processing while obscuring the original sensitive data, enabling external computation without direct exposure of proprietary information.
2Measurement precision
If raw input data is sent directly to the machine learning model, then processing accuracy is maintained, but information leakage risk increases
Solution Approach 1:
The patent applies preliminary action by performing pre-processing of the input data locally before sending it to the external machine learning model. This pre-processing step generates multiple variants of the input data through transformations such as adding noise, changing formats, or creating augmented versions. By preparing the data in advance with these protective transformations, the system maintains processing accuracy while reducing information leakage risk.
Solution Approach 2:
The patent utilizes parameter changes by transforming the input data into multiple variants with different parameters (e.g., adding random noise with different magnitudes, applying different transformation parameters). These parameter variations allow the machine learning model to process the data effectively while making it difficult to reconstruct the original sensitive information from any single variant, thus balancing accuracy and privacy.
3Loss of information
If multiple input variants are generated through pre-processing, then data privacy is enhanced, but processing complexity increases
Solution Approach 1:
The patent applies the copying principle by creating multiple simplified variants of the input data instead of implementing a single complex privacy-preserving mechanism. Each variant is a copy transformed through relatively simple operations (e.g., adding noise, format conversion). This approach enhances privacy through multiple copies while keeping the complexity of individual processing steps manageable, as each variant can be processed independently through standard machine learning inference.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
According to various embodiments, a method for processing data by means of a machine learning model is described, comprising receiving input data for the machine learning model, generating multiple input variants from the input data by differently pre-processing the input data in multiple different ways by one or more predetermined pre-processing methods, obtaining, for each input variant, at least one respective output variant by supplying the input variant to the machine learning model and determining output data from the obtained output variants.