Estimation of Microphone Clusters in Acoustic Sensor Networks using Unsupervised Federated Learning

Abstract

In this paper we present a privacy-aware method for estimating source-dominated microphone clusters in the context of acoustic sensor networks (ASNs). The approach is based on clustered federated learning which we adapt to unsupervised scenarios by employing a light-weight autoencoder model. The model is further optimized for training on very scarce data. In order to best harness the benefits of clustered microphone nodes in ASN applications, a method for the computation of cluster membership values is introduced. We validate the performance of the proposed approach using clustering-based measures and a network-wide classification task.

Unsupervised Clustered Federated Learning in Complex Multi-source Acoustic Environments

“…For both 4SA and 2SL scenes, we can indeed observe the aforementioned effects for the first four and two columns, respectively. When compared to [15] and [19], which both use unfurnished shoebox rooms, it can be observed that the results generated by the improved version of unsupervised CFL proposed in this work show great potential, especially for the 2SL scene which clearly outperforms [19].…”

“…Moreover, hardclustering without the possibility of generating membership values has been used. We have recently addressed some of the aforementioned aspects and successfully explored the application of an unsupervised CFL clustering scheme in ASNs [19]. However, the latter work is also limited to a shoebox room with only two simultaneously active sources and, consequently, only two clusters.…”

“…the amount of training data clients can train on, potentially leading to overfitting issues. For this reason, [19] proposes pre-training an autoencoder h on more extensive data and then freezing all its layers except the bottleneck layer. The latter can be re-initialized with random values and further adapted with much smaller amounts of data by individual clients.…”

“…As observed in fuzzy clustering approaches [15], determining cluster membership values (MVs) offers vital benefits to ASN-based applications. While standard CFL provides only hard clustering results, we have previously extended CFL to also compute MVs [19]. We now refine this method in order to cope with more than two sources and two clusters.…”

“…In order to evaluate clustering performance in relation to previous work [19] and state-of-the-art implementations [15], we extend the normalized cluster-to-source distance (CTS) dsz cx of cluster c x to source s z [15] to multiple sources and clusters,…”