ASR Language Model Rescoring for Memorization Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Modern automatic speech recognition (ASR) systems that leverage language models (LMs) unintentionally memorize sensitive user-specific text data during training, creating a privacy vulnerability that can be exploited by hackers.
Innovation Solution
Insert canary text samples into the training corpus and use a text-to-speech system to generate synthetic speech utterances, rescoring these with an external language model to detect memorization by analyzing the word error rate (WER), and apply per-sample gradient clipping to mitigate this memorization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a language model is used to disambiguate acoustically similar transcriptions in ASR systems, then transcription accuracy is improved, but the system unintentionally memorizes sensitive text data creating privacy vulnerabilities
Solution Approach 1:
The patent applies preliminary action by inserting canary text samples into the training corpus before training the language model. This proactive measure allows the system to detect memorization of sensitive data through canary tokens that trigger detection mechanisms, enabling privacy protection before actual sensitive data is compromised.
Solution Approach 2:
The patent uses canary text samples as an intermediary mechanism to detect memorization. These canary samples act as a mediator between the training process and security monitoring, allowing the system to indirectly detect whether sensitive data has been memorized without directly exposing the sensitive data itself.
2Reliability
If canary text samples are inserted into the training corpus and used for detection, then privacy security is improved, but the training process and evaluation complexity increase
Solution Approach 1:
The canary text samples serve multiple functions: they are inserted into the training corpus to train the language model, and simultaneously serve as detection triggers to monitor for memorization. This multi-functionality reduces the need for separate detection mechanisms and simplifies the overall system architecture.
Solution Approach 2:
The language model itself performs the detection function by generating transcriptions that reveal whether canary text samples have been memorized. The model's normal operation during evaluation serves the dual purpose of both transcription and security verification, eliminating the need for separate detection processes.
3Reliability
If gradient clipping is applied to mitigate memorization, then privacy protection is improved, but training efficiency and convergence speed may be reduced
Solution Approach 1:
The patent applies gradient clipping selectively to gradients originating from canary text samples rather than all training data. This partial application of gradient clipping mitigates memorization of sensitive data while minimizing the impact on overall training efficiency, as the clipping is applied only to a subset of the training corpus.
Data Source
AI summary
A method includes inserting a set of canary text samples into a corpus of training text samples and training an external language model on the corpus of training text samples and the set of canary text samples inserted into the corpus of training text samples. For each canary text sample, the method also includes generating a corresponding synthetic speech utterance and generating an initial transcription for the corresponding synthetic speech utterance. The method also includes rescoring the initial transcription generated for each corresponding synthetic speech utterance using the external language model. The method also includes determining a word error rate (WER) of the external language model based on the rescored initial transcriptions and the canary text samples and detecting memorization of the canary text samples by the external language model based on the WER of the external language model.