Privacy concerns in federated learning have attracted considerable attention recently. In centralized networks, it has been observed that even without directly exchanging raw training data, the exchange of other so-called intermediate parameters such as weights/gradients can still potentially reveal private information. However, there has been relatively less research conducted on privacy concerns in decentralized networks.

In this report, we analyze privacy leakage in optimization-based decentralized federated learning, which adopts generally distributed optimization schemes such as ADMM or PDMM in federated learning. By combining local updates with global aggregations, it was proved that optimization-based approaches are more advantageous compared to the traditional average consensus-based approaches, especially in scenarios where the data at the nodes are not independent and identically distributed (non-IID).

We further extend the privacy bound in distributed optimization to the decentralized learning framework. Different from the fact in the centralized learning framework the leaked information is the local gradients of each individual participant at all rounds, we find that in decentralized cases the leaked information is the difference of the local gradients within a certain time interval. Motivated by the gradient inversion in centralized networks, we then design a homogeneous attack to iteratively optimize dummy data whose gradient differences are close to the true revealed gradient differences. Though the gradient difference information still brings privacy concerns, we show that it is more challenging for adversaries to reconstruct private data using the difference of gradients than using the gradients themselves in the centralized case.

To deal with the privacy attack, we propose several potential defense strategies such as early stopping, inexact update and quantization etc. The main advantage of these approaches is that they introduce error/noise/distortion into decentralized federated learning for protecting private information from being revealed to others without affecting the training accuracy. In addition, we also show that the larger the batchsize is, the more difficult for the adversary to reconstruct the private information.

In recent years, the large increase in connected devices and the data that is collected by these devices has caused a heightened interest in distributed processing. Many practical distributed networks are of heterogeneous nature. Because of this, algorithms operating within these networks need to be simple, robust against network dynamics and resource efficient. Additionally, if privacy preservation methods are properly implemented, they add to the power of distributed processing by making it possible to leverage the data of many different users, without infringing the privacy of the individuals involved.
In this study we focus on the primal-dual method of multipliers (PDMM), which is a promising dis- tributed optimisation algorithm that seems to be suitable for distributed optimisation in heterogeneous networks. Most theoretical work that can be found in existing literature focuses on synchronous ver- sions of PDMM. However, in heterogeneous networks, asynchronous algorithms are favourable over synchronous algorithms. So far, simulation results have indicated that asynchronous PDMM converges and can even converge in the presence of transmission losses.
In this work we analyse the properties of stochastic PDMM, which is a general framework that can model variations of PDMM such as asynchronous PDMM and PDMM with transmission losses. We build upon previous empirical results of PDMM and formulate theoretical proofs to substantiate these results. After defining stochastic PDMM and proving its convergence, we compare a number of PDMM variations that have been mentioned throughout the literature. Lastly, we derive a lower bound for the variance of the auxiliary variable in the context of stochastic PDMM, assuming uniform updating probabilities. This lower bound indicates that subspace based privacy preservation is applicable to certain instances of stochastic PDMM, like asynchronous PDMM.
The main result of this work is a theoretical proof that shows that stochastic PDMM converges almost surely if the updating probabilities of each auxiliary variable are nonzero. Two important conclusions that follow from this proof are the almost sure convergence of asynchronous PDMM and unicast PDMM with transmission losses. Another useful result is the fact that subspace based privacy preservation is effective when using asynchronous PDMM.

Acoustic room geometry estimation is often performed in ad hoc settings, i.e., using multiple microphones and sources distributed around the room, or assuming control over the excitation signals. To facilitate practical applications, we propose a fully convolutional network (FCN) that localizes reflective surfaces under the relaxed assumptions that (i) a compact array of only two microphones is available, (ii) emitter and receivers are not synchronized, and (iii), both the excitation signals and the impulse responses of the enclosures are unknown.
Our FCN is designed to extract spectral and temporal patterns from stereo recordings, aggregate the temporal information over time-frames, and predict the likelihood of virtual sources corresponding to reflective surfaces at specific locations. Whereas most source localization algorithms are limited to direction-of-arrival (DOA) estimation, the proposed method jointly estimates distances and DOAs. Numerical experiments confirm that the network is able to generalize to mismatched microphone array sizes, sensor directivity patterns, or audio signal types, while highlighting front-back ambiguity as a prominent source of uncertainty. When a single reflective surface is present, up to 80% of the sources are detected, while this figure approaches 50% in rectangular rooms.
Further tests on real-world recordings report similar accuracy as with artificially reverberated speech signals, validating the generalization capabilities of the framework.

Environmental sound identification and recognition aim to detect sound events within an audio clip. This technology is useful in many real-world applications such as security systems, smart vehicle navigation and surveillance of noise pollution, etc. Research on this topic has received increased attention in recent years. Performance is increasing rapidly as a result of deep learning methods. In this project, our goal is to realize urban sound classification using several neural network models. We select log-Mel spectrogram as the audio representation and use two types of neural networks to perform the classification task. The first is the convolutional neural network (CNN), which is the most straightforward and widely used method for a classification problem. The second type of network is autoencoder based models. This type of model includes the variational autoencoder (VAE), beta-VAE and bounded information rate variational autoencoder (BIR-VAE). The encoders of these systems extract a low dimensionality representation. The classification is then performed on this so-called latent representation. Our experiments assess the performances of different models by evaluation metrics. The results show that CNN is the most promising classifier in our case, autoencoder-based models can successfully reconstruct the log-Mel spectrogram and the latent features learned by encoders are meaningful as classification can be achieved.

Consensus problem has been a topic of interest for many research areas allowing multiple agents to reach an agreement through local information exchange. The explicit share of the state variables, however, may cause privacy issues due to the confidentiality of the initial values. In this work, asynchronous privacy-preserving consensus average algorithms are proposed which enables the agents to reach the exact average of their initial values while preserving the privacy of them. The research aims to reduce the convergence time and computational complexity compared to the cryptographic solutions. Three methods are proposed and compared. The state decomposition and noise-obfuscation methods preserve the privacy of the initial values given that the semi-honest adversary is not able to listen to one of the neighboring nodes of the targeted node. The hybrid state decomposition approach proposes a way to overcome this assumption by using a minimum number of encryption operations. The initial values are shown to be private against an eavesdropper who is able to tap all communication channels as well as a semi-honest adversary in the system. It has been shown in all proposed approaches that as the noise variance goes to infinity, the adversary does not have any range to estimate the initial value. The noise obfuscation technique futures the same convergence rate as the standard averaging approach while providing a linear increase in the variance of the adversary's estimate with the increasing noise variance. On the other hand, the state decomposition technique futures a lower convergence rate compared to the standard averaging approach while providing an exponential increase in the variance of the adversary's estimate with the increasing noise variance. By optimizing the algorithm, it has been shown that the same convergence rate as the standard randomized gossip can be obtained. The state decomposition approach requires the addition of noise for a bounded amount whereas, the noise obfuscation method requires the addition of noise at each iteration. All three approaches, converges faster than a fully cryptographic approach while promising statistical security guarantees.

Biometrics authentication has been very useful and necessary nowadays due to the great developments in technology and the transaction of huge amounts of sensitive data on a daily basis. Traditionally, access to some data or service is achieved by means of some documents or a password. However, these methods are not very convenient. Alternatively, typical biometric systems can be employed that use fingerprint, iris, voice, face recognition or a combination of them. This project focuses on the task of face recognition from still images and investigates how different algorithms for face verification perform under various adverse conditions modelled by blur, salt-and-pepper noise and changes in illumination. Conventional pattern recognition algorithms are first presented. Pixel intensities, Gabor features, Local Binary Patterns (LBP) and 2D-DCT coefficients are considered as features while for classification the nearest neighbor (NNC), nearest mean (NMC), SVM classifiers, and Likelihood Ratio Tests (LRT) with Gaussian Mixture Models (GMM) are examined. Out of all these methods, Gabor features combined with the linear SVM classifier are shown to produce best results across all degradations giving an average Equal Error Rate (EER) of 0:97% using the ORL face dataset. Then, emphasis is placed on deep learning and Convolutional Neural Networks (CNN). Specically, VGG-Face with triplet loss training for face verification is suggested. VGG-Face achieves an average EER of 2:63% when both test images of a query image pair are drawn from the same degradation conditions and an average EER of 3:80% when only one image in the given pair is degraded and the other one is derived from the clean ORL dataset. We also experimented with the extracted VGG-Face features and NNC, linear SVM and Gaussian SVM and it is seen that a linear SVM gives an average EER of 1:10% by macro-averaging the Detection Error Tradeoff (DET) curves.

The particle filter (PF) algorithm is appropriate to solve the problem of speaker tracking in a reverberant and noisy environment using distributed pairwise microphone networks. First, complete the tracking task based on PF algorithm in centralized manner, a processing center is required to collect the signal from all microphones to carry out the PF processing. The computation complexity and time consumption of the particle filter algorithm are relatively high, mainly because of the large number of particles exploited in the filtering process since the effectiveness and accuracy of the particle filter particularly rely on the sample set size. However, almost all the existing particle filtering algorithms exploit the fixed number of particles, especially in the field of acoustic source tracking. To deal with this matter, Kullback-Leibler distance (KLD) sampling method was utilized as an adaptation technique to adjust the sample size instead of setting fixed number. Two approaches based on particle filter algorithm for tracking speaker in distributed way are proposed. Compared to the centralized scheme, each microphone pair in the distributed network executes the local PF individually and exchanges local weights or posterior parameters among neighboring nodes to efficiently achieve the global estimate of the sound source position. Finally, simulation experiments demonstrate these two methods are feasible to track the speaker in distributed microphone networks with a variable number of particles.

The response of a sound system in a room primarily varies with the room itself, the position of the loudspeakers and the listening position. The room boundaries cause reflections of the sound that can lead to undesired effects such as echoes, resonances or reverberation. Knowledge of the location of these large reflecting surfaces can help to better estimate the sound field behavior inside the room. This work focuses on exploiting the inherent information present in echoes measured by microphones to infer the location of nearby reflecting surfaces. The investigated application uses a loudspeaker to emit a known signal and record the resulting sound field with a co-located microphone array. A signal model is proposed which provides a relationship between reflector locations and measured microphone signals. The locations of reflections are estimated by fitting a sparse set of modeled reflections with measurements. We present two novelties with respect to prior art. First, the method is end-to-end where from raw microphones measurements it outputs an estimate of the location of reflectors. For the case of a compact uniform circular microphone array, the symmetry can be exploited to reduce the computational complexity of the inference process. Secondly, the model is extended to include a loudspeaker model that is aware of the inherent directivity pattern of the loudspeaker. The performance of the proposed localization method is compared in simulation to the existing state-of-the-art localization methods. An experimental study with real world measurements was also conducted to investigate the performance of the model.

The thesis project is aimed at designing an unobtrusive method to find the obstruction location and severity for patients who are not diagnosed with Obstructive sleep apnea, during their non-sedated sleep using simple recording devices within uncontrolled environment. Similar to speech generation, which is enabled by opening and closing of the vocal cords, the sound of snoring consists of a series of impulses caused by the rapid obstruction and reopening of the upper airway. By exploiting this similarity, we try to explain snoring sound production using synthesis techniques that has been applied to speech. A trial has been conducted to gain information on efficacy of different commercially available devices that are used to alleviate snoring problem. Sleeping sounds from this trial has been analyzed to find a method to find the effective device for each participant. Similar to speech analysis, Iterative Adaptive Inverse Filtering(IAIF) method has been used to find excitation flow and airway tract transfer functions of snoring sounds during the inhalation period. A model has been built using the Acoustic Tube Theory with the purpose of reflecting the physical realities of snoring sound production. Resulting transfer function spectra from acoustic tube modeling and IAIF has been compared using a gain-independent Itakura Spectral Distance Measure. It has been found that Linear Prediction is not suitable to directly determine the cross-sectional area of the upper airway from snoring sounds. An alternative method, using transmission line model with acoustic tube modeling has been found to produce transfer functions that are spectrally similar to transfer functions obtained from source filter models.

A phase estimation algorithm is presented to estimate the phase of a recurring pattern in a nonstationary signal. The signal is modelled by a template signal that represents one revolution of the recurring pattern, and that the frequency of this pattern can change at any time with no assumptions about local stationarity. The algorithm uses a constraint maximum likelihood estimator (MLE) to estimate the phase of the recurring pattern in the time series. Using the dynamic programming techniques from the dynamic time warping (DTW) algorithm, the solution is found in an efficient manner. The algorithm is applied to the digitization of meter readings from analog consumption meters.

As of today, analog consumption meters are still widely used to measure the consumption of gas, electricity and water. Often, smart home appliance use a simple reflective photosensor located on a rotating part of the meter to obtain information about the state of the consumption meter. The algorithm presented in this thesis accurately estimates the phase of the repeating pattern that occurs in the sensor observation when the meter rotates. Using this estimate, the signal of the photosensor can be converted to an estimate of the total resource consumption and consumption rate.

The algorithm improves in accuracy over conventional methods based on peak detection, and is shown to work in cases where the peak detection methods fails. Examples of this are signals where there is no distinctive peak in the signal or a signal where the recurring pattern is reversed. Furthermore, a template compression scheme is proposed that is used to decrease the computational complexity of the algorithm. Different time series compression methods are applied to the algorithm and evaluated on their performance.

Estimating a room geometry using multiple microphones rises an echoes labeling problem. Two recent methods called the graph-based and the subspace-greedy methods have shown their capability in solving this problem. The graph-based method attains a good accuracy but suffers in maintaining the computational cost when the number of microphones is larger than 7. On the other hand, the subspace-greedy method provides suboptimal accuracy with much lower computational time. Here we construct the hybrid combination methods using those two baseline methods by interchanging their intermediate steps: the refinement step and the source localization step. To assess their practicability in a real-life application such as virtual reality games and robot navigation, the performance of these hybrid methods were tested against the close microphones arrangement on the sphere's surface. However, this new microphones' constellation brings up a low dimensional problem. To deal with this matter, we use the weighted least squares as the source localization procedure. Finally, experiments on synthetic squared distance data demonstrate the feasibility of all hybrid methods for estimating the room geometry with centimeter precision within seconds.

Authentication is becoming an increasingly important application in the connected world and is driven by the growing use of mobile and IoT devices that use an increasing number of applications that require transactions of sensitive data. Security usually relies on passwords and/or two-factor authentication which are too intrusive for daily use. Biometric solutions such as fingerprints, voice, iris and retina are a good alternative to overcome previous problems. In this project an audio-visual identity verification is presented, where the use of multiple modes that can already be captured from most IoT devices (microphone and camera) make authentication robust to adverse conditions. End-factor analysis (i-vectors) with cosine distance is implemented as the main classification algorithm which takes into account variations within and between speakers. Mel Frequencies Cepstrum Coefficients (MFCC) are used as audio features, 2D-DCT coefficients of a single snapshot and Motion Vectors (MV) of the lips are extracted for visual features. Improvements combining different modes are shown using VidTimit dataset where the proposed algorithm achieves 0.7% of Half Total Error (HTER) in the test set outperforming single modes audio and visual by 9.5% and 6.4%, respectively.

In this thesis report, the design of the estimation of internal delays within speakers and microphones will be covered. The estimation is done via an iterational algorithm which converges to the different delays. The design choice of the estimation of those delays follows from an initial attempt to improve synchronization of the Bosch DICENTIS microphone units. With an unit being a system is placed on the desk of an attendee of a conference, consisting of a microphone and a speaker. After coming to the conclusion that the current level of synchronization already sufficed the focus was shifted towards the estimation of the internal delays.

The choice of algorithm that was used for the estimation is covered in the in the state of the art analysis of the current internal delay estimation techniques. The subsystem receives pre-determined times between units and uses these to estimate the internal delays. This estimation is done with a random initialization of the delays (within reasonable margins for the delays). After which this estimation converges towards the real values by minimizing a Frobenius norm between the rank three approximation and the received times. This is elaborated on in the Estimating Internal Delays section. The algorithm can also make use of a regularization term which decreases the time required for the estimation of the delays. The results of the algorithm are discussed in the Implementation section, which consists of a number of MATLAB simulations using the implemented algorithm. Using the results, a conclusion is drawn for the viability of the solution after which a recommendation of future work is given.

In this report, the design of a subsystem within a localization system for the Bosch DICENTIS wireless conference system will be presented. The localization system will function by means of using acoustic Time Difference Of Arrival (TDOA) measurements to determine the location of each unit connected to the DICENTIS conference system. By unit, the system on the desk of each attendee in the conference, that contains a microphone and speaker is meant. The task of the subsystem presented is to estimate propagation times of transmitted signals between speakers of each unit in the conference system and the microphones of each unit.

The design choice for the type of localization method implemented is based on the gathered information from an initial literature study, the hardware specifications of the Bosch DICENTIS system and the demands for the localization system that were imposed by Bosch. The subsystem will function by transmitting a set of pseudo-random codes, modulated using a type of Frequency Shift Keying (FSK), where two On Off Keying (OOK) signals, modulated at different frequencies, are superimposed. The received and demodulated pseudo-random codes are then correlated with multiple different peak detectors that will correlate with multiple different sets of the transmitted string of pseudo-random codes to gain a higher robustness for the estimated propagation times and a higher accuracy for these estimates. Results show that the the use of multiple different sets of transmitted codes indeed improves the propagation time estimation. The overall system as presented, concerning accuracy and robustness, meets the requirements made by Bosch. However, in future work, optimalization of the system with regard to computation time is required.

Distributed optimization has been an
extensively studied field for years. Recent developments in the area of sensors makes it possible to create networks consisting of a large number of nodes. The focus of this thesis will be optimizing distributed problems over a decentralized network. These distributed optimization schemes operate in an iterative matter as follows. First each node performs some local computations, after which the data is transmitted to its neighbours. The purpose of this study is to investigate the effects of approximating these local computations inexactly on the convergence of distributed optimization schemes. Although we consider many optimization schemes in general, the primal-dual method of multipliers (PDMM) is used during the simulations. Therefore we start off by deriving the inexact iteration for PDMM which shows how the inexactness propagates through the iterates. This derivation also suggests that the inexactness depends on the optimization constant, which was verified during the simulations. After that, the convergence of distributed optimization schemes is analyzed by making use of monotone operator theory to investigate under
which conditions convergence will be reached. This convergence analysis has two main results. It firstly shows that distributed optimization schemes converge to a fixed point if the error is summable and secondly
that an error has less influence as iterations pass. Thereafter simulations are presented that suggest that the inexactness affects how far the algorithm converges, thus what the remaining error is when convergence is
reached. Decreasing the error when convergence is reached causes the inexact PDMM iteration to resume converging at the rate of the standard PDMM algorithm. These observations holds in synchronous as well as asynchronous operation. Introducing packet loss only influences the convergence rate of the inexact PDMM iteration.

Physical activity recognition through wearables has enabled the development of novel applications in healthcare. Most of the existing studies focus on predicting activities using wearable sensors, either in a controlled or uncontrolled environment. However, there is not a clear distinction between these two environments. Hence, this thesis aimed to answer the research question “How accurately can we classify physical activity based on wearable accelerometers placed on the wrist and chest in a controlled and in a free-living environment?".

For the data collection phase, two experiments were conducted in the working environment of imec. 40 participants were recruited and were asked to participate in the Controlled and Free-Living Study. The subjects wore two imec wearables, a wrist-worn and chest-worn accelerometer sensor and performed everyday activities. These activities include sitting, dynamic sitting, lying with face up and face down, lying to the left and right, standing, dynamic standing, walking upstairs, walking downstairs, walking, running, and cycling. The Controlled Study showed that most of these activities could be detected accurately using accelerometer data from both sensors with 91.83% F1-score. Similarly, the combination of these two sensors achieved the best performance for the Free-Living Study with 86.98% F1-score. Finally, this work proved that between the two environments a correlation could be possible only for the activity cycling. Consequently, this research concludes that the activity recognition should be explicitly investigated in free-living environments, focusing on real-time activity detection.

Clock synchronization among the nodes of a wireless acoustic sensor network (WASN) is a significant issue that affects the performance of multi-channel noise reduction schemes. Since independent sensors are utilized, each accompanied by its internal clock, clock offsets are inevitable, even if the mismatch in the sampling frequencies is negligible. In this thesis, clock offsets are mathematically modeled and the problem of multi-channel linear filtering for speech enhancement is addressed through signal subspace methods. For this purpose, the generalized eigenvalue decomposition (GEVD) of the cross-power spectral density (CPSD) matrices of the noise and target speech processes is capitalized. Beamformers based on this technique are proved to be invariant to sensor clock offsets when used in a blind manner, exploiting only network measurements. This result is confirmed through experiments in a simulated environment.

Human interaction with a smart speaker involves often distant automatic speech recognition (ASR). However, ASR is a rather cumbersome task at significantly high levels of noise. Most of commercial smart speakers in order to achieve high ASR accuracy they tend to reduce the playback signal once the preset keyword is detected. In an effort to dispose this function from the smart speaker, in this thesis a speech enhancement technique is considered in the front-end of the ASR system aiming at the suppression of the dominant noise component in the degraded speech signal. Having a priori knowledge on the playback signal renders adaptive filtering a well-suited speech technique. Therefore, the class of least mean squares (LMS) algorithms is studied and assessed. Among other techniques of this class the transform domain LMS (TDLMS), due to its inherent signal decorrelation properties, is shown to achieve the best performance in terms of noise suppression and improved speech intelligibility as well as word error rate. The results of this study correspond to a set of simulation incorporating real impulse responses measured in both an anechoic and a reverberant environment.

Recently, the effects of quantization on the Primal-Dual Method of Multipliers were studied.
In this thesis, we have used this method as an example to further investigate the effects of quantization on distributed optimization schemes in a much broader sense. Using monotone operator theory, the effect of quantization on all distributed optimization algorithms that can be cast as a monotone operator was researched for two different problem subclasses. The averaging problem was used as an example of a quadratic problem, while the Gaussian channel capacity problem was an example of the non-linear problem subclass. A fixed bit rate quantizer was used in combination with a dynamic cell width, to analyse the robustness of distributed optimization schemes against quantization effects. In particular, we have shown that for practical implementations it is possible to incorporate fixed bit rate quantization with dynamic cell width in a distributed optimization algorithm without loss of performance for both problem classes.

Packet communication applications cannot guarantee correct delivery of every packet. Congestions and interferences in the network lead to lost packets. However, real-time applications require timely delivery of data or information and always tolerate packet loss to achieve this aim. When some speech packets are lost, packet loss concealment (PLC) is used to replace the missing speech.

In this thesis, after investigating packet loss characteristics in realistic wireless networks and problems in existing PLC algorithms, we propose a new PLC scheme named adaptive PLC, which is composed of three algorithms: odd-even interpolation, waveform similarity matching and silence substitution. Adaptive PLC adjusts the algorithm to use depending on loss situations. Odd-even interpolation recovers the loss by interpolating odd or even samples in a packet. Waveform similarity matching estimates waveform segments from correctly received or already recovered packets. Silence substitution just fills in the missing part by zeros.

The adaptive PLC achieves improvements in speech quality relative to each single PLC algorithm and other existing PLC algorithms.