Federated Learning (FL) is a decentralized machine learning approach that provides a privacy-friendly way of training models by keeping the datasets of participating parties private. Some challenges FL faces are the lack of incentives to encourage participation in the learning process, as well as preventing potential cyber attacks that tamper with the model. Blockchain is an available solution that provides the means to implement incentives to encourage participation and issue penalties to disincentivize malicious behavior. Hence, recent developments introduced blockchain-enabled FL (BCFL) designs for various applications. However, one obstacle that slows down the widespread adoption of this technology is the high latency of blockchain networks due to its laborious consensus protocols. In this paper, we propose StateFL, a revised BCFL architecture that uses state channels (a blockchain scaling solution) in order to ease the load on the blockchain by reducing the number of on-chain transactions, improving the system’s latency, and minimizing transaction fees as a result. State channels are governed by smart contracts and enable two parties to exchange information and assets off-chain unless disputes occur. Either channel party can dispute the state of the channel if suspicious behavior is observed. In that case, the dispute is settled on the blockchain. We evaluate StateFL in a series of experiments to establish latency improvements, identify bottlenecks, and quantify the impact of disputes on channel latency and transaction fees. The findings show that the higher the number of FL rounds, the more StateFL outperforms the baseline BCFL, with the exception of a very low number of rounds. In realistic FL scenarios, the rounds are in the order of hundreds making StateFL a solid contender even if disputes do occasionally occur. The bottleneck of StateFL is the channel setup and closure which require extensive interaction with the blockchain.

Foremost among the challenges of the Bitcoin blockchain is the scalability bottleneck. To address this issue, the Lightning Network, a payment channel network, was created. Lightning is a payment channel network that is source-routed and uses onion routing, like Tor. However, unlike Tor, the routing path is determined by optimizing a cost function, which uses public information. In this paper, a timing attack is evaluated by simulation of the Lightning Network. The goal of the attack is for an adversary that is part of a payment to determine the destination of the payment. The influence colluding adversaries have on the performance of the attack is evaluated. Three types of colluding adversaries (closest to destination, farthest from destination, average of all adversaries) are compared to a simple adversary that guesses the next hop in the payment is the destination of the payment. It is found that the adversary closest to the destination performs the best. Furthermore, shadow routing, a mitigation against these types of attacks, is evaluated against this attack. It is found that shadow routing does not have a significant impact on the performance of this attack.

Payment channels allow parties to utilize the blockchain to send transactions for a cheaper fee. Previous work has analyzed to which degree a party can profit by facilitating the transaction process. The aim is to increase the usability of the network and to be rewarded for providing this service. However, previous work focuses on maximizing the reward of the individual player in isolation, a model that we aim to expand. That is why in this work we extend the action space to allow other parties to act and react, and observe the impact this has on the rewards of the player that would otherwise act in isolation.
Testing existing placement strategies by performing channel placement games, we can assess the difference in the reward that indicates the potential loss that competition may cause when operating in the Bitcoin Lightning Network.
Furthermore, we have developed a new strategy that is able to improve the performance in the multi-actor model.

Privacy is a human right, yet, people’s behavior on the web is constantly tracked. Tor, an anonymity network, is an effective defence against tracking. However, Tor’s multiplexing of logically independent data streams into a single TCP connection causes issues. Tor with QUIC has been implemented as an alternative with better performance but it has not been studied whether and by how much QUIC increases the vulnerability to timing-based attacks.
The most threatening attacks are website fingerprinting attacks, which can track a Tor user by only controlling the guard node, first of the relays that forward traffic in Tor. In this work, Tor with QUIC is evaluated against website fingerprinting attacks with various levels of defences active. Without defences, Tor is vulnerable to website fingerprinting for both TCP and QUIC but the attacks are more effective on QUIC. On the positive side, defences against website fingerprinting remain effective for QUIC in that they decrease the effectiveness of the attack by a
similar fraction as for TCP.

Blockchain technology has proven to be a promising solution for decentralized systems in various industries. At the core of a blockchain system is the peer-to-peer (P2P) overlay, which facilitates communication be- tween parties in the blockchain system. Recently, there is increasing evidence that this P2P overlay plays a major role in limited transaction throughput, limited scalability, and security threats in a blockchain system. In this work, it is explained that this bottleneck is caused by the fact that currently deployed P2P overlays cannot fully capture the complexity of a blockchain system. We propose a novel approach to address these issues by introducing a novel overlay architecture. Our approach, Sovereign Domains P2P Overlays, allows for nodes in a single network to belong to multiple heterogeneous groups (called domains), each with their own set of protocols tailored to the characteristics and needs of the nodes in that domain. To demonstrate the effectiveness of the proposed overlay architecture, two novel node discovery protocols, FedKad and SovKad, were designed and implemented. The simulations conducted in a custom-built simulator show that SovKad outperforms node discovery in a Structured P2P Overlay (Kademlia) and node discovery in a Federated P2P Overlay (FedKad), providing evidence of the potential of the Sovereign Domains P2P Overlay architecture in blockchain systems.

The set of regression and integration tests at many modern software companies is huge. It is difficult to run all tests after each code change, so the tests are often run for batches of code changes by different developers, late in the release cycle. This has developers waiting for tests longer than desirable and makes it difficult to link a failed test to a specific code change. To increase developer productivity a subset of high-priority tests can be selected to be run directly after a code change. Tests can furthermore be reordered so that tests reveal faults faster. This field is called test case selection and prioritization (TCS&P). On the flip side the large number of historical test executions can be leveraged to predict test failures. This can be done by training machine learning models on the historical execution data and metrics about the software repository at which these tests were executed. The prediction of failure is then used as the basis for test selection and the confidence of the prediction can be used for prioritization, with higher risk tests being run earlier resulting in tests failing faster on average. This technique is named predictive TCS&P (PTCS&P). This work describes an attempt at PTCS&P by using the historical test executions of Adyen, a financial technology platform. The resulting model has a recall of 95.4% and can be used for three use cases: skipping the tests with a low probability of reveal- ing a fault, which misses 0.6% of faults and saves 34.4% wall clock time, selecting only tests with a high probability of revealing a fault, which saves 94% of time and reveals 39% of faults, and prioritizing based on the chance of failure, which allows for partial left shifting of these tests, which reduces the average time to first fault revealed by 95%.

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.

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.

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.

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.

Begun in 2020, Polkadot is one of the largest blockchains in market capitalization and development. However, privacy on the Polkadot network has yet to be one of the key focus points. Especially unlinkability between the user’s IP address and Polkadot address is essential. Without this unlinkability, users are vulnerable to targeted ads, manipulation, blackmail, reputational damage, financial loss, physical harm, discrimination, and more. This thesis investigates the viability of Tor or a VPN with Polkadot as external privacy-enhancing tools to hide the user’s IP address, as users aiming to achieve unlinkability cannot easily change the Polkadot code.

To analyze the viability, we set up a measurement study to examine the performance of a Polkadot full node behind Tor or a VPN. We investigated, among other things, the latency, throughput, and the number of discovered and connected peers to determine the performance of three Polkadot full nodes located in London, Seoul, and North California. Furthermore, we did a security analysis to determine any vulnerabilities that could emerge from using Polkadot with either of the network environments. And we investigated in-depth the susceptibility of the Polkadot node to an Eclipse attack, as previous research has shown that Bitcoin with Tor was vulnerable to an Eclipse attack.

Our results show that a Polkadot node with Tor has considerably high latency and cannot maintain long-lasting connections. The short connection time decreases the time to perform an Eclipse attack on a Polkadot node from a couple of months and weeks for the normal and VPN environment to potentially six days or less for the Tor environment. We calculated the cost of running an Eclipse attack to be approximately €482 per week. The Polkadot node behind the VPN does perform considerably better. The Polkadot node in London, behind the VPN located in Frankfurt, performed similarly in terms of latency to the Polkadot node in a normal network environment. However, the Polkadot nodes in both
the Tor and VPN environment have only outgoing connections. If too many nodes ran behind one of these environments, fewer peers would be able to establish connections with one another, resulting in network partitions or network failure.

This study emphasizes the importance of unlinkability between a Polkadot user’s address and IP. However, using Tor or a VPN as privacy-enhancing tools could impact the security of the Polkadot node and the whole Polkadot network. So users should avoid using Tor with Polkadot and carefully consider the tradeoff between privacy and security when using a VPN. The security issues mentioned in this thesis should be further investigated and tested. Furthermore, a default solution built into the Polkadot source code should be investigated.

In the past 8 years, Bitcoin has dominated the cryptocurrency markets and drawn attention from academia, developers and legislators alike. Bitcoin has been praised for its impact on decentralizing trust and currencies but also criticized for its volatility and energy-inefficient consensus mechanism. To improve its limitations, in 2016, payment channels and payment channel networks were introduced in the form of the Lightning Network. Payment channels allow for so-called off-chain transactions that, in case of dispute, can be published to an existing cryptocurrency blockchain, like Bitcoin, for arbitration. After its introduction, the concept of payment channels was quickly adopted by many cryptocurrency users. However, although payment channels remove the need for many on-chain transactions, some still remain. An on-chain transaction is required for the opening and closing of a channel. This happens during the initial setup between two users but is also required if one of the users runs out of balance on their side of the channel. The latter is a common occurrence as transactions are often unidirectional, say between a customer and a merchant. To limit the amount of closing and opening on-chain transactions required, a user can start or take part in a rebalancing. A rebalancing is a process with the aim of bringing a channel to a balance as desired by its owners. The state-of-the-art existing protocol to carry out a rebalancing is called Revive, which is a distributed protocol using leader election and a linear program to calculate the optimal rebalancing between its participants. Although effective, the protocol provides little privacy to its participants. We, therefore, designed a new, privacy-preserving peer-to-peer rebalancing protocol. Alongside it, we also introduce an accompanying participant discovery protocol that allows users in a network to find other users interested in running a distributed algorithm. We show that both protocols are secure and that our rebalancing protocol provides more privacy than Revive, at the cost of a suboptimal result and an increased message and time complexity. Finally, we compare our rebalancing protocol and Revive using a payment channel network simulator that simulates transactions taking place during the rebalancing. Using this simulation, we show that both protocols have a negative effect on the payment channel network as they lock the to-be-rebalanced channels while they are executing. We, therefore, conclude that an ideal rebalancing protocol should both be privacy-preserving and concurrent, and propose ideas to achieve this in future research.

There is an everlasting arms race between censoring bodies and those in its grip. When the censor is employing increasingly sophisticated techniques to digitally monitor and restrict those in its scope, equally sophisticated means to circumvent the digital repression come forward. Those suffering under digital censorship are using nifty ways to escape the censor's grip. Digitally restrictive regimes occasionally still allow access to blockchain applications such as cryptocurrencies, albeit in limited form. Blockchains often enjoy a global nature, but traditional and well established cryptocurrencies such as Bitcoin often do not have high performance. Second layer solutions, also known as off-chain solutions, offer a network of payment channels where transactions can be completed in peer-to-peer fashion with little interaction with the slow blockchain. In this thesis we investigate how Bitcoin's off-chain solution, the Lightning Network, can be employed to circumvent digital censorship. We introduce Periscope, a protocol that allows for tunneling of internet traffic between two hosts over a stream of micro-transactions embedded with data. Next, we thoroughly analyse its security and the network's suitability from a theoretical perspective. Following this, an empirical evaluation study is done to get an understanding of the performance of the protocol under various circumstances. We hope that Periscope serves as an additional mean to access the free internet to those in need.

Users of anonymity networks face differential treatment and sometimes get blocked by websites, it is currently unclear how common this blocking is. This research aims to provide an overview of how common this blocking is while utilizing the AN.ON anonymity network. The analysis is accomplished by utilizing automated web scraping and processing to recognise and classify blocks by comparing them to a control connection. This process and software can be used and extended to analyze and compare any two connections. The scope is limited to the one thousand most popular websites according to the Alexa rating. Different kinds of blocks were identified and automatically recognised in processing, though manual verification is still required. Evidence is found and presented that there is a significant amount of blocking, occurring on approximately 23% of the analyzed domains. There is also a significant difference in blocking between using different cascades.

Censorship and privacy issues have led people to use VPNs when accessing the internet. These VPNs not only try to protect their user but they are also associated with criminality and cyber attacks. Because of this, websites have started to resort to blacklisting the IP addresses that are used by the VPNs, thus blocking both genuine and malicious users. This forces users to sacrifice privacy for accessibility. This paper provides a method on how to measure the amount of blocking that VPN users experience and to be able to determine what type of blocking is occuring. This method is then used in an experiment using a web crawler where nodes from ProtonVPN are used to measure the amount of blocking that occurs while browsing the internet’s most popular websites. This experiment shows that on average 1.12% of the domains perform some type of blocking directed towards the VPN user and that the majority of this blocking consists of a total block, which means that the user is entirely excluded from any use of the website. Next to this it is shown that not all VPN nodes show the same amount of blocking and that there was no large difference in blocking found between days while using the same VPN node. It also shows that the categories which perform the most blocking are Business, Online Shopping and News.

Tor is an anonymity network used by a vast number of users in order to protect their privacy on the internet. It should not come as a surprise that this service is also used for abuse such as Denial of service attacks and other malicious activities because of the anonymity it provides. For protecting themselves from this abuse, websites block Tor in various ways. We investigate the extent and frequency of this kind of blocking by requesting the Alexa top 1000 websites with and without Tor with the objective of highlighting the differential treatment observed by privacy-minded users. We build upon existing studies by using diverse metrics to measure discrimination and by extending our search to three sub pages of websites for detecting sophisticated blocking. We find at least 25.8% of the Alexa top 1000 websites discriminating on the home page against Tor users as opposed to 20.03% observed in previous studies. This number rises to 31.7% after including the three sub pages. We also discover new types of blocks such as Tor users being served old or different versions of websites. We categorize the blocked websites and find that Online Shopping and Finance/ Banking categories discriminate most against Tor while Social Networking sites and Search Engines discriminate the least.

Anonymity networks, such as The Invisible Internet Project, commonly known as I2P, enable privacy aware users to stay anonymous on the Internet and provide secure methods of communication, as well as multi-layered encryption. Despite the many innocent reasons users opt for online anonymity, these particular networks are censored at times, as they are associated with criminal activity. The goal of this paper is to measure to what extent I2P network users are being blocked by popular websites, and not, however, by governments or internet service providers. To establish this, we developed a web crawler which compares the responses to HTTP(S) GET requests sent anonymously, via I2P, and non-anonymously. Our results are based on the analysis of the received HTTP status codes, and on screenshots of the requested websites, to assess content blocking. This experiment shows that I2P users suffer from some form of blocking in 10.09% of cases. However, it should be noted that I2P faces certain bandwidth limitations and traffic congestion at the outproxy. This is a result of the fact that I2P was not designed with the intent of being a proxy to the Internet, but rather a self sustaining peer-to-peer network.

There are many valid reasons for someone to choose to stay anonymous online, not least of which is the fact that online privacy is a human right. However, discrimination against users of anonymity networks from web-servers and content distribution networks on the grounds of defense against malicious users often means that genuine users are faced with excessive challenge-response tests and differentiated content, or even blocked altogether. This study has investigated the extent to which users of Mullvad VPN are blocked when accessing popular websites and it has also explored the nature of these blocks. No statistically significant difference was found when requesting only home pages from 3,000 domains, but this changed when classifying 1,000 domains and considering content beyond the home page. This impact on the user's experience is also reflected on the categories of website that most engage in blocking, with some essential services such as health and government presenting high blocking ratios. Given that more discerning ways of preventing access from malicious users without affecting genuine ones exist, this generalised blocking of Mullvad VPN users is unjustifiable.

The Lightning Network is a second layer payment protocol built on top of Bitcoin, which is scalable and has reduced transaction fees. It does so by eliminating the need to broadcast every transaction to the whole network. When one user wants to send a payment to another, the routing protocol generates a path between them that is always fast and cost efficient. The low degree of randomness in the existing routing protocols during path selection allows an adversary to compromise the anonymity of the sender and recipient.
In this work, we propose a new routing algorithm that is less predictable when creating a transaction path. We show that this increases the anonymity of the users of the Lightning Network by creating an attack on the new routing protocol. The attacker tries to identify the potential source and recipient of a transaction. Our results suggest that there is a trade-off between the offered anonymity and transaction fees; it is not possible to get higher anonymity at no cost by designing a non-deterministic routing algorithm.