Homomorphic encryption is a type of encryption which allows computations to be performed on encrypted data, without the need for the data to be decrypted first. Recent works have shown that homomorphic encryption can be used to train encrypted machine learning models on untrusted hardware. Federated learning enables distributed training of machine learning models on remote devices with their own private datasets. Existing federated learning techniques focus on protecting the privacy of the remote data, but not on protecting the content of the model being trained. Homomorphic encryption can be used with federated learning to protect the model. Homomorphic encryption is very computationally expensive, however, it has been shown that GPU (Graphics Processing Unit) acceleration can be used to decrease its required computation time. Modern GPUs, including those in mobile devices, can be used for general purpose computing. GPUs are well suited to data-parallel tasks in which one operation is applied to many items. Many of the computations involved in homomorphic encryption are well suited to a data-parallel approach. This research investigates the use and implementation of GPU accelerated homomorphic encryption on mobile devices and examines the potential for its use in federated learning tasks. There are many types of homomorphic encryption, some of which are better than others for certain types of computation. For instance, HEAAN (Homomorphic Encryption for Arithmetic of Approximate Numbers), also known as CKKS (the initials of the authors of HEAAN), is well suited to computations involving vectorized fixed-point values. We implement and benchmark HEAAN for mobile GPUs. We anticipate that the practically achievable model depth will be very limited even with GPU acceleration, but that it will be significantly better than CPU implementations. We also anticipate that memory requirements will be a significant limiting factor for practical model depth.

Deep embedded clustering (DEC) utilizes deep neural networks (DNN) in order to learn feature representations using an autoencoder which is optimized for clustering. This is done by integrating a clustering loss using Kullback-Leiber divergence (KL divergence). Autoencoder models have been successfully applied to many types of data in order to enable unsupervised representation learning. Recurrent neural networks and long-short term memory (LSTM) networks have been utilized in learning representations of audio data. In many cases, convolutional autoencoder algorithms (CAE) have been used in processing audio data, in order to extract their feature representations. However, for the purpose of clustering similar learned embedded features from audio data, there has not yet been an integration of DEC in the LSTM autoencoders. This research project focuses on implementing DEC for audio signals. For this purpose, we have integrated a clustering loss using KL divergence into a LSTM autoencoder. Mel Spectrograms of the audio data are then extracted. This time-series data is fed into the network. We have evaluated our model performance using enormous data sets of audio signals collected from deep and shallow water hydrophones. With the decrease of hardware costs, stationary hydrophones are increasingly deployed in the marine environment to record animal vocalizations amidst ocean noise over an extended period of time. Bioacoustic data collected in this way is an important and practical source to study vocally active marine species and can make an important contribution to ecosystem monitoring. However, a main challenge of this data is the lack of annotation which many supervised neural network models rely on to learn to distinguish between noise and marine animal vocalizations. In contrast to the previous works done in this field, our approach is designed for unsupervised representation learning, allowing us to use a large volume of unlabeled data.