Description
Federated Learning (FL) provides a promising alternative to centralized machine learning for anomalous sound detection. Instead of transmitting raw training data to a central instance where a model is trained, FL leverages the computational power on sensor devices: a model is updated locally and only the weight updates are sent to a fusion center, which then generates a global model. This way, the data remains on-device. Nevertheless, studies have shown that some information about data is still present in weight updates, indicating risks of privacy leakage. This thesis aims to develop an add-on for FL that conceals critical information of weight updates that are being sent to a potentially untrustworthy fusion center.

Task
Develop and evaluate a neural network based method to conceal critical information in weight updates while keeping a satisfying anomalous sound detection performance.

Prerequisites

• Strong Python skills
• Solid knowledge of acoustics and signal processing
• Interest in artificial intelligence and deep neural networks