Analyzing privacy leakage in distributed algorithms is challenging as it is difficult to track the information leakage across different iterations. In this paper, we take the first step to conduct a theoretical analysis of the information flow in distributed optimization ensuring that gradients at every iteration remain concealed from others. Specifically, we derive a privacy bound on the minimum information available to the adversary when the optimization accuracy is kept uncompromised. By analyzing the derived bound we show that the privacy leakage depends heavily on the optimization objectives, especially the linearity of the system. To understand how the bound affects privacy, we consider two canonical federated learning (FL) applications including linear regression and neural networks. We find that in the first case protecting the gradients alone is inadequate for protecting the private data, as the established bound potentially exposes all sensitive information. For more complex applications such as neural networks, protecting the gradients can provide certain privacy advantages as it will be more difficult for the adversary to infer the private inputs. Numerical validations are presented to consolidate our theoretical results.

@en

We are surrounded by all kinds of sounds at all times. What we hear varies with the physical environment and our position. Room impulse responses (RIRs) characterize the effect of the environment on a sound produced by a source. A first goal of this dissertation is to analyze RIRs and investigate how to extract environmental information from RIRs. Immersive digital environments, such as virtual reality (VR) and augmented reality (AR), play an increasingly important role in society. Spatial audio, aiming to give listeners a 3D audio experience, is vital to immersive digital environments. Omnidirectional RIRs do not provide explicit spatial information for room acoustics applications. As a result, the description and reproduction of the sound field is of great importance for spatial audio. Specifically, we consider higher order ambisonics, which is the prevalent method to represent the sound field around a listener. The synthesis of ambisonics signals is a second goal of this dissertation.... @en

Mapping a room impulse response (RIR) to its Ambisonics representation is not always feasible. However, by adding a weak assumption (i.e., the existence of at least two perpendicular walls in the environment), the Ambisonics representation is restricted to be one of a finite set, with known transformations between the set entries. This makes mapping the omnidirectional RIR to the Ambisonics RIR (ARIR) possible. The authors solve the mapping problem with a convolutional neural network and multi-task variational autoencoder. The room is assumed to be rectangular. The proposed method is based on the image source method with frequency-independent reflection coefficients exclusively. The authors focus on the early part of RIRs, where the directional information lies. This method requires only a single RIR. Generalizing to the real world, measurements can obviate the need for specialized hardware for Ambisonics measurement. The proposed method can achieve an SNR of 17.62 dB on estimated first-order ARIRs and 16.15 dB on estimated third-order ARIRs.

@en

We describe a new method to estimate the geometry of a room and reflection coefficients given room impulse responses. The method utilizes convolutional neural networks to estimate the room geometry and multilayer perceptrons to estimate the reflection coefficients. The mean square error is used as the loss function. In contrast to existing methods, we do not require the knowledge of the relative positions of sources and receivers in the room. The method can be used with only a single RIR between one source and one receiver. For simulated environments, the proposed estimation method can achieve an average of 0.04 m accuracy for each dimension in room geometry estimation and 0.09 accuracy in reflection coefficients. For real-world environments, the room geometry estimation method achieves an accuracy of an average of 0.065 m for each dimension.

@en

We investigate what information about a room is necessary to integrate a new source into an existing scenario. In particular, we consider the effects of the reflection order, the order of ambisonics signals and reverberation time. We conducted a series of listening tests and used the control variates method to determine the quantitative relevance of the selected attributes. In terms of integration and accurate localisation, at least third order ambisonics description of a source, is required for integration of that source. In addition, a finite number of early reflections can perform equally well to a full room impulse response when a new source is integrated into an existing scenario. However, the room impulse response with only the correct reverberation time is not sufficient.

@en

Indoor localisation is an important research topic with several possible applications. For example, knowing a user's location can be used as navigation aid in hospitals and malls, or for better targeted marketing. In this paper we consider the case where the environment of interest is equipped with several receivers (with known location) from which time-difference-of-arrival (TDOA) measurements are obtained and used to localise the source. We will present a distributed algorithm for localising the source. More specifically, we experimentally show that the distributed algorithm, which only uses time-of-arrival (TOA) measurements obtained from neighbouring receivers to calculate the TDOAs, performs as well as a centralised solution that has access to all TOA measurements in the network. In addition, we propose a method for discarding erroneous TOA measurements which considerably improves the performance in noisy and reverberant environments.

@en