Optical flow estimation with event cameras encompasses two primary algorithm classes: model-based and learning-based methods. Model-based approaches, do not require any training data while learning-based approaches utilize datasets of events to train neural networks. To effectively apply these algorithms, it's essential to understand their respective strengths and weaknesses.
This study compares model-based and learning-based optical flow estimation methods using event cameras, aiming to provide guidance for real-world applications. We evaluated these methods on the MVSEC and DSEC datasets, focusing on their accuracy and runtime. Our findings indicate that model-based methods excel on the MVSEC dataset, characterized by small motions, while learning-based approaches perform better on the more dynamic DSEC dataset. To investigate potential overfitting of learning-based methods to DSEC, we retrained the IDNet and TMA models on the BlinkFlow dataset. The retrained models demonstrated competitive accuracy, surpassing model-based methods which indicates that learning-based models perform better on datasets like DSEC even when not able to overfit. Finally, our analysis on runtime showed that model-based methods achieve real-time performance on CPUs and learning-based methods require a GPU to run in real-time.

Computer vision tasks have shown to benefit greatly from both developments in deep learning networks, and the emergence of event cameras. Deep networks can require a large amount of training data, which is not readily available for event cameras, specifically for optical flow estimation. The need for simulating this data in a realistic, physics-driven manner is therefore crucial. This paper compares the state of the art event camera simulators on different criteria, including event timestamp modeling, performance under low illumination, bandwidth simulation, computation speed and various types of noise simulation. We also summarize the shortcomings of some commonly used optical flow event datasets. For generating high-quality, realistic events, The V2E and DVS-Voltmeter simulators have shown to produce the most accurate data.

Event cameras are bio-inspired sensors with high dynamic range, high temporal resolution, and low power consumption. These features enable precise motion detection even in challenging lighting conditions and fast-changing scenes, rendering them well-suited for optical flow estimation. However, event camera output is sparse and unstructured, making it challenging to process. Transformer architectures have shown to be effective in capturing long-term temporal dependencies and processing sparse input, hence they might be better suited to processing this output by leveraging the fine time granularity inherent to event camera data.
We introduce E-GMFlow, an approach for event-based optical flow inspired by the recent success in terms of accuracy of transformer-based models for frame-based optical flow. We explore the effect of temporal details on the accuracy of this transformer architecture by changing the number of temporal bins in which events are discretized. We observe that the increase in the number of temporal bins generally causes higher accuracy and comment on the limitations of this study.

The aim of this paper is to complete the gap in the knowledge and experiment using as little as only the heart rate of some subjects to manage to successfully authorise them in some supposed system. The focus will be on the Gaussian Mixture model and the One Class Support Vector Machine, both outlier detectors, because most of the past research was focused on supervised models. Using these two, this paper will experiment with recognising intruders with models trained to distinguish one authorised person and multiple authorised persons. In the first case, multiple data processing methods and hyperparameters will be tested together and compared. In the second case, the goal will be to use and modify the best-found parameters of the first case to train models that are able to detect multiple persons as authorised. This time, there are two methods that this paper is going to look into and compare their performance: training one single model to detect all the subjects and training multiple models, one per subject. The most notable results are with one authorised person, with a score of 0.936, with two authorised, 0.88, and with 12 authorised, 0.713, when using the area under the curve metric.

Outlier detection is an essential part of modern systems. It is used to detect anomalies in behaviour or performance of systems or subjects, such as fall detection in smartwatches or voltage irregularity detection in batteries. This provides early indications of something of potential problems.

A part of outlier detection that is not often analysed is the performance of algorithms in environments with data from only one subject, versus environments with data from multiple subjects. This paper aims to answer the questions regarding the performance of Gaussian Mixture Models (GMM) and DBSCAN in these different environments. This paper focuses on time series data collected from consumer-grade wearables like smartwatches. In this paper, the outliers are defined manually, as the used data set did not contain predefined outliers. This research considers both outliers defined within the subject data, and the use of other subjects as outliers.

Results from this paper indicate that the amount of subjects in the environment is not the sole factor in the performance of these algorithms. Rather, it is a combination of the amount of subjects in the environment and the type of outlier to be detected. Results show that a GMM has difficulty distinguishing subjects that are similar when using another subject as outlier data. On average, DBSCAN outperforms a GMM in almost all cases, and DBSCAN is a lot more consistent in its performance than a GMM.

Heart rate data and other data collected by consumer-grade wearable devices can give away quite useful information about the user. It can for example be used by machine learning algorithms such as Deep Neural Networks (DNN) to learn patterns about cardiovascular disease and fitness, or be used for identification. Heart rate patterns can also change quickly within the span of several months, which could make older heart rate data less useful when training a DNN. This paper shows that the DNN did indeed perform significantly worse when trying to identify people on older data compared to recent data. The accuracy calculated from the test set was 63.64% when trained on the most recently available training data, in comparison to 33.88% when trained on the least recent data which was more than 200 days older. When changing the recentness of training data only for a single user, there was also always an improvement in the accuracy of the model to identify that particular person. The accuracy to identify all users however did not necessarily increase, and sometimes even decreased. Using more data for training still outperforms using a smaller amount of samples of more recent data by slight margins, showing the trade-off between the recentness of data and the amount of data used for training. However, if fast training times are required, taking the most recent data windows can still lead to a similar performance as when training on all available data.

Advancements in the precision and accuracy of consumer-grade wearables, such as a Fitbit, have enabled the identification and therefore authentication of individuals based on their emitted heart frequencies using these wrist-worn devices. With this type of authentication, a password is essentially sent out every second. This makes it a perfect form of authentication in fields where constant authentication is crucial. However, not much is known about how different types of cardiac diagnosis (e.g. fit or obese) influence the accuracy of this type of authentication. In this paper, it will be shown how different types of cardiac diagnosis affect the accuracy of Deep Neural Networks to identify individuals by heart rate. This study is done with data obtained from 14 subjects, having different types of cardiac diagnosis. A deep neural network consisting of multiple convolutional layers is being used to conduct the experiments. It has been shown that subjects with a paroxysmal atrial fibrillation diagnosis improve the accuracy the most, compared to the reference (normal healthy) subjects. On the other hand, (very) fit subjects decrease the accuracy the most. Heart failure and obese subjects have a similar accuracy compared to reference subjects.

In recent years, with the rapid expansion of IoT (Internet of Things) devices, more and more research and commercial projects have focused on various application areas of IoT. Signify, as a leading player in the smart home industry, has been deeply involved in this field for many years, particularly focusing on smart lighting for smart homes and providing consumers with a whole-house smart lighting solution (Hue System). The Hue System consists of several IoT devices, such as the Hue Bridge, Hue Bulbs, and Hue Accessories. This research paper specifically targets the IoT field and aims to reduce the firmware update cycle of the Hue Bridge.

The Hue Bridge, which serves as a central device in the Philips Hue internet-connected lighting system, connects users with other Hue devices. However, the Hue Bridge faces the challenge of inefficient firmware updates, which require validation engineers to wait for 6-8 weeks to ensure firmware reliability. To address this issue, this paper proposes a virtual system solution that improves the virtualization procedure of the Hue Bridge devices and utilizes Kubernetes for large-scale deployment to accelerate the generation of diagnostic data. Furthermore, a Use Case Model is established based on users' daily data, and a model based on Frequent Pattern Mining is applied to simulate users' daily behaviors in the Kubernetes Deployment.

To validate our virtual system, we designed validation experiments from multiple perspectives, including validation of the use case model and automated feedback. Our validation results demonstrate that this system enables more efficient and convenient acquisition of automated feedback (issues/bugs), while significantly enhancing the generation of diagnostic data in the firmware update cycle. Moreover, it offers advantages such as high availability, convenience, and cost-effectiveness in deployment. This research provides valuable references and insights for firmware update-related studies in the IoT domain

Badnets are a type of backdoor attack that aims at manipulating the behavior of Convolutional Neural Networks. The training is modified such that when certain triggers appear in the inputs the CNN is going to behave accordingly. In this paper, we apply this type of backdoor attack to a regression task on gaze estimation. We examine different triggers to discover which of them lead to better performance and thus infer which trigger aspects one can take the most advantage from. It turns out that placing frames around and drawing multiple lines across the images are the most effective for the training of Badnets.

Recent years have seen an increasing interest in stablecoins from major corporate and governmental parties. The European Central Bank is investigating the possibility of introducing its own Central Bank Digital Currency. The desired features of such a currency are under discussion. One such feature is offline spending: the ability to use the currency without an internet connection, like cash. This thesis describes a token-based transaction prototype and its implementation on the Kotlin-IPv8 protocol stack. The prototype allows funds to be spent in an offline setting and provides retroactive fraud detection. The prototype is not intended for deployment but instead serves as a trial for building digital currencies on Kotlin-IPv8. The included performance analysis demonstrates that various facets of Kotlin-IPv8 perform suboptimally, of which most notably its UDP data throughput.

Model extraction attacks are attacks which generate a substitute model of a targeted victim neural network. It is possible to perform these attacks without a preexisting dataset, but doing so requires a very high number of queries to be sent to the victim model. This is otfen in the realm of several million queries. The more difficult the dataset, the more queries required to gain an accurate substitute model. Through each state-of-the-art model extraction algorithm, one thing that is not thoroughly optimised are the hyperparameters of the models, and optimizing them has been found to have a strong impact on accuracy of the substitute model. To attempt to reduce the number of queries required, research has been done to find the effects of optimizing hyperparameters for both MNIST and fashionMNIST datasets. This is done through grid search and random search. The results show that proper hyperparameter tuning can reduce the number of queries required to perform model stealing if they are not already optimized. Examples include requiring 125000 + queries to achieve 95% accuracy for the MNIST dataset with some hyperparameter combinations to only requiring 15000.

Machine Learning (ML) models are vulnerable to adversarial samples — human imperceptible changes to regular input to elicit wrong output on a given model. Plenty of adversarial attacks assume an attacker has access to the underlying model or access to the data used to train the model. Instead, in this paper we focus on the effects the data distributions has on the transferability of adversarial samples under a ``black-box'' scenario. We assume an attacker has to train a separate model (the ``substitute model'') and generate adversaries using this independent model. The substitute models are trained with different data distributions: symmetric, cross-section or completely disjoint data to the one used to train the target model. The results demonstrate that an attacker only needs semantically similar data to execute an effective attack using a substitute model and well-known gradient based adversarial generation techniques. Under ideal attack scenarios, target model accuracies can drop below 50\%. Furthermore, our research shows that generating adversarial images from an ensemble increases average attack success.

Adversarial training and its variants have become the standard defense against adversarial attacks - perturbed inputs designed to fool the model. Boosting techniques such as Adaboost have been successful for binary classification problems, however, there is limited research in the application of them for providing adversarial robustness. In this work, we explore the question: How can AdaBoost ensemble learning provide adversarial robustness to white-box attacks when the "weak" learners are neural networks that do adversarial training? We design an extension of AdaBoost to support adversarial training in a multiclass setting, and name it Adven. To answer the question, we systematically study the effect of six variables of Adven’s training procedure on adversarial robustness. From a theoretical standpoint, our experiments show that known characteristics from adversarial training and ensemble learning apply in the combined context. Empirically, we demonstrate that an Adven ensemble is more robust than a single learner in every scenario. Using the best found values of the six tested variables, we derive an Adven ensemble that can defend against 91.88% of PGD attacks and obtain 96.72% accuracy on the MNIST dataset.

A machine learning classifier can be tricked us- ing adversarial attacks, attacks that alter images slightly to make the target model misclassify the image. To create adversarial attacks on black-box classifiers, a substitute model can be created us- ing model stealing. The research question this re- port address is the topic of using model stealing while minimizing the amount of querying the sub- stitute model needs to train. The solution used in this report is a variant of the ActiveThief algo- rithm that makes use of active learning to deter- mine which data is being queried. The paper exper- iments with different subset selection strategies to find the most informative data points. Also, a seed- ing algorithm based on clustering is explored and finally, a stopping criterion for the ActiveThief al- gorithm is proposed. These variations are evaluated on their accuracy and the number of queries they take to achieve that accuracy. This paper shows cluster seeding is an alternative to random seeding in ActiveThief. This paper also presents different subset selection strategies that outperform the ran- dom sampling strategy. Finally, a stopping criterion based on entropy is introduced that halts the algo- rithm when an uncertainty threshold is reached.

In recent years, there has been a great deal of studies about the optimisation of generating adversarial examples for Deep Neural Networks (DNNs) in a black-box environment. The use of gradient-based techniques to get the adversarial images in a minimal amount of input-output correspondence with the attacked model has been extensively studied. However, existing studies have not been discussing the effect of different gradient estimation techniques coherently. In this paper, a new one-point residual estimate is compared to the known two-point estimates. The findings in this paper show that the one-point residual estimate is not a viable option to decrease the number of queries to the attacked model. The accuracy of the attacks with the use of an one-point residual estimate maintains the same for weaker models. For stronger models, there is a slight decrease in accuracy at identical distortion levels. All estimates are tested on different PGD attacks on the MNIST and F-MNIST datasets using a 3-layer convolutional network.

Natural Language Interfaces for Databases (NLIDBs) offer a way for users to reason about data. It does not require the user to know the data structure, its relations, or familiarity with a query language like SQL. It only requires the use of Natural Language. This thesis focuses on a subset of NLIDBs, namely those with 'plain English' sentences as input and SQL queries as output.

Study 1 recruits participants from multiple origins (i.e. academia, a crowdsourcing platform, banking industry) without selection based on their query language capabilities. Next, participants are segmented based on query language capabilities to distinguish between non-experts and experts. A common way to retrieve information from databases is by using SQL. Thus knowledge of SQL is assumed to be a proxy for participants' skill level (i.e. SQL proficient, non-SQL proficient). We create an approach that uses an automated near semantic equivalence evaluation for user-generated queries against a predefined gold-standard SQL query and thus segment participants. We find that 70 out of 242 participants are identified as SQL proficient. To differentiate between the segmentations, we define 42 requirements often implemented for NLIDB systems, from which both segmentations pick a selection as their preferred requirements. We are unable to find statistically significant differences between the segmentations' preferences. However, exploratory findings reveal the importance of origin, namely the banking industry, which prefers explanation over answer accuracy, different from other segmentations.

Study 2 is inspired by the exploratory findings of Study 1 and uses requirements from Study 1 to create an application that tests two conditions, one with an explanation by using color-coding (i.e. to show the relations between the natural language question asked and the models' output columns) and another without. NLIDBs make it hard for users to verify if the answer provided by its model is correct. Therefore, Study 2 uses these two conditions above to test if color-coding improves performance for the participants. Our findings suggest that color-coding only improves performance for non-aggregate selection queries with multiple columns.

Recent works explain the DNN models that perform image classification tasks following the "attribution, human-in-the-loop, extraction" workflow. However, little work has looked into such an approach for explaining DNN models for language or multimodal tasks. To address this gap, we propose a framework that explains and assesses the model utilizing both the categorical/numerical features and the text while optimizing the "attribution, human-in-the-loop, extraction" workflow. In particular, our framework deals with limited human resources, especially when domain experts are required for human-in-the-loop tasks. It provides insight regarding which set of data should the human-in-the-loop tasks be brought in. We share the results of applying this framework to a multimodal transformer that performs text classification tasks for compliance detection in the financial context.