Hearing Device DNN Copy Detection Without Reverse Engineering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge of detecting a copied deep neural network (DNN) in hearing devices is significant due to the difficulty in distinguishing between authentic and unauthorized copies without compromising the device's functionality or requiring reverse engineering, especially in cases where DNNs are used for noise reduction.
Innovation Solution
A system and method involving a challenge signal to train the DNN to respond with a predetermined output, allowing for the detection of an authentic DNN by analyzing its response to the signal without affecting its primary noise reduction function.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a challenge signal method is used to detect copied DNNs, then detection capability is improved, but device complexity increases
Solution Approach 1:
The detection system is segmented into separate components: a challenge signal generator external to the hearing device, a transmitter to deliver the challenge signal, and a receiver to capture the DNN response. This segmentation allows detection functionality to be added without complicating the internal structure of the hearing device itself, resolving the contradiction between detection capability and device complexity.
Solution Approach 2:
A challenge signal acts as an intermediary between the detection system and the DNN. Instead of directly analyzing the DNN's internal structure or requiring reverse engineering, the system uses an external challenge signal to elicit a response that reveals the DNN's authenticity. This intermediary approach enables detection while maintaining simplicity in the hearing device architecture.
2Measurement precision
If reverse engineering is used to detect copied DNNs, then detection capability is improved, but loss of information increases
Solution Approach 1:
The detection method extracts only the necessary response information from the DNN without requiring access to its internal structure, training data, or proprietary algorithms. By taking out only the challenge signal response, the system achieves detection capability while preserving all other information within the DNN, including noise reduction capabilities and proprietary intellectual property.
Solution Approach 2:
The challenge signal and its response serve as an intermediary mechanism that enables detection without direct access to the DNN's internal information. This approach prevents loss of information by avoiding reverse engineering while still providing sufficient data to determine DNN authenticity through the external challenge-response interaction.
3Measurement precision
If the DNN is retrained with challenge signals, then detection capability is improved, but productivity decreases
Solution Approach 1:
The DNN is preliminarily trained during the development phase to recognize specific challenge signals and produce predetermined responses. This preliminary action is performed once during DNN creation, after which the trained DNN can be deployed in hearing devices without requiring repeated retraining. The challenge signal detection capability is built into the DNN from the start, enabling ongoing detection without impacting productivity.
Solution Approach 2:
The detection capability is copied into the DNN during its initial training process. Instead of requiring continuous retraining or external processing, the challenge signal recognition and response generation are embedded as inherent functions of the DNN itself. This copying of detection functionality into the DNN structure enables efficient operation without reducing productivity.
Data Source
AI summary
System and methods are presented for detecting an authentic copy of a deep neural network in a device without reverse engineering the device. For example, the device can be a hearing device. A method is presented that includes providing audio signals to a deep neural network located in a hearing device, receiving signals from the deep neural network that are triggered by the transmitted audio signals, and determining whether the received signals were the expected signals. Specifically, the system can determine that an unauthorized entity embedded an unauthorized copy of the deep neural network in a hearing device.