Obfuscating Training Data for Audio Analysis Privacy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Audio analysis systems, such as speech recognition and language identification, rely heavily on large training datasets, which can be costly and time-consuming to source, and often contain confidential information that cannot be shared without consent, posing a challenge in maintaining data privacy.
Innovation Solution
The method involves obfuscating training data by randomizing and reorganizing annotated feature vectors generated from audio and text transcripts, ensuring that the audio analysis system cannot determine the content or subject matter of the original data, while still utilizing the data for model training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If large training datasets are used to improve accuracy of audio analysis systems, then the accuracy and representativeness of the system improves, but data privacy and confidentiality are compromised
Solution Approach 1:
The patent extracts only the essential acoustic features and state information from the training data while removing or obscuring the original audio content and text transcripts. This allows the system to retain the useful training information needed for accuracy while eliminating the confidential content that poses privacy risks.
Solution Approach 2:
The patent creates obfuscated copies of the training data that preserve the acoustic patterns and linguistic features needed for model training, but replace the original confidential audio and text with randomized or synthetic representations that cannot be traced back to the source material.
2Productivity
If existing confidential conversations and transcripts are used for training, then the cost and time of data collection is reduced, but consent and privacy requirements cannot be met
Solution Approach 1:
The patent introduces an obfuscation process as an intermediary between the confidential training data and the audio analysis system. This intermediary transforms the data into an intermediate representation that maintains training value while removing privacy concerns, enabling the use of existing confidential data without direct consent.
Solution Approach 2:
The patent changes the parameters of the training data by transforming audio signals into acoustic feature vectors and text into state sequences, altering the data representation from recognizable content to abstract numerical representations that preserve statistical patterns but eliminate identifiable information.
3Manufacturing precision
If original training data is shared with audio analysis systems, then model training quality improves, but data security and confidentiality controls are violated
Solution Approach 1:
The patent converts the harmful aspect of confidential data (its sensitivity and identifiability) into a benefit by using the obfuscation process to create training data that is equally valuable for model training but inherently secure. The very process that removes identifying information preserves the acoustic and linguistic patterns needed for high-quality training.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Examples disclosed herein involve obfuscating training data. An example method includes computing a sequence of acoustic features from audio data of training data, the training data comprising the audio data and a corresponding text transcript; mapping the acoustic features to acoustic model states to generate annotated feature vectors, the annotated feature vectors comprising the acoustic features and corresponding context from the text transcript; and providing a randomized sequence of the annotated feature vectors as obfuscated training data to an audio analysis system.