Current Research Areas

Anonymity and Censorship Resistance

Access to important information, such as news or medical information, is limited in large parts of the world due to governmental censorship. Furthermore, access if allowed is frequently tracked, which can have serious consequences. For instance, accessing information about abortion can lead to investigations and potential arrests if abortion is prohibited in a region.

It is hence important to provide solutions that enable people to access critical information without having to fear for their safety. As the situations regarding freedom of speech and privacy vary widely, we require different solutions for different scenarios. Scenarios with fewer limitations allow for more lightweight and efficient solutions whereas solutions for highly restricted scenarios usually affect performance more due to the higher level of stealthiness they have to provide.

Currently, our research focusses on the following aspects:

• Ad-hoc networks as an alternative means of communication during internet shutdowns
• Censorship-resistant publication using anonymous P2P networks
• Anonymous off-chain payments for distributed ledgers

If you are interested in a Bachelor or Master thesis regarding ad-hoc networks and their role during internet shutdowns, please contact Shen Yu (y.shen-5 AT tudelft.nl). For the other topics, contact Stefanie Roos (stefanie.roos AT cs.rptu.de)

Privacy and Security in Distributed Machine Learning

Centralized machine learning relies on obtaining data from data providers/sources. Optimally, there should be multiple data providers to have a good data diversity. However, the data may be confidential and/or sensitive, e.g., medical data. Thus, data providers may be unwilling to or legally prohibited from sharing the data. We explore methods that enable parties to train a model locally without the data leaving their premises.

However, the trained models still reveal information about the original data. Furthermore, as the data is not shared, it is non-trivial to detect whether data or models have been manipulated. Our research here is two-fold: We design attacks on privacy and security of distributed machine learning to identify vulnerabilities. We then modify the learning process to counteract these attacks.

Currently, we are looking into the following distributed machine learning scenarios:

• Federated Learning and Multi-discriminator Generative Adversarial Models: One centralised party selects clients that train and aggregates their results
• Multi-Federator Learning: Multiple parties propose learning tasks for similar data and then select clients from a shared pool
• Decentralized Learning: Learning one of multiple tasks without a central party guiding the process

If you are interested in a Bachelor or Master thesis in this area, please contact Jiyue Huang (J.Huang-4 AT tudelft.nl) or Stefanie Roos (stefanie.roos AT cs.rptu.de).

Distributed Ledger Scalability

Distributed ledgers enable traceability without a trusted third party. They have first been introduced to realize currencies and transactions without banks. They are furthermore useful when a set of parties that do not mutually trust each other want to build a shared (digital) infrastructure.

Typically, all information needs to be broadcast to all parties in the distributed ledger in order to reach consensus on the ledger state. However, that causes high latencies and high communication overhead. In our research, we explore protocols that resolve some issues locally and only require global consensus when necessary.

Currently, we are looking into the following blockchain scalability solutions:

• Payment Channel Networks (e.g., Lightning): parties open channels using shared deposits and can conduct transactions locally using these deposits
• Multi-Chain Systems (e.g., Polkadot): It is possible to have different chains for different applications/purposes and connect them to a main chain while maintaining the security guarantees of the main chain.

If you are interested in a Bachelor or Master thesis in this area, please contact Stefanie Roos (stefanie.roos AT cs.rptu.de).