The following projects are available for Master and Bachelor students. They are performed in close collaboration with an experienced member of the lab. Apply for a project by sending an email to the contact mentioned for the project.
You may also suggest new projects, ideally close enough to our ongoing, or previously completed projects. In that case, you will have to convince Anne-Marie Kermarrec that it is worthwhile, of reasonable scope, and that someone in the lab can mentor you!
Projects available for Spring/Summer 2026.
Free-rider detection in decentralized learning models using watermarks
Master’s thesis or MSc semester project
Contact: Jade Bourrée ([email protected])
Algorithmic auditing consists of independently evaluating whether an automated system behaves in a fair and unbiased way. To do so, an auditor needs data: ideally, a representative sample of the platform’s users and their experiences. In practice, however, the data available online is far from representative. On platforms like TripAdvisor or TikTok, users who feel treated unfairly are much more likely to speak up than those who are satisfied, leaving the auditor with a dataset that is mostly made up of complaints and missing a large silent majority.
This project investigates whether techniques from positive and unlabeled (PU) learning, a branch of machine learning designed to work with incomplete and one-sided data, can help auditors make better use of what they have. The goal is to explore how a simple surrogate model, built on top of PU learning methods, could allow an auditor to compensate for this imbalance and draw more reliable conclusions about the platform’s behavior.
To contribute effectively to this project, we highly value:
- Strong ML fundamentals and proficiency in ML implementation
- Strong mathematical foundation
[1] AI Auditing: the broken bus on the road to AI accountability, SaTML 2024.[2] Analysis of learning from positive and unlabelled data, NeurIPS 2014.
Robust Decentralized Learning via Model Fragmentation
Master’s thesis or MSc semester project
Contact: Sayan Biswas ([email protected]) or Martijn de Vos ([email protected])
Decentralised learning (DL) is emerging as a compelling alternative to centralised and federated training, enabling collaborative model optimisation without a central server or data sharing. A recent paradigm introduces model fragmentation, whereby nodes decompose locally trained models into smaller fragments and disseminate them independently across the network. This approach has demonstrated strong privacy benefits [1] and improved robustness to asynchrony by mitigating straggler effects [2], while maintaining high accuracy.
However, DL remains vulnerable to active adversaries and Byzantine nodes, whose malicious updates can significantly degrade model utility.
This project aims to investigate whether model fragmentation can also serve as a defence mechanism against such attacks. Specifically, we study whether distributing “low-intensity” poisoned updates (i.e., smaller fragments) across many peers leads to less accuracy degradation than propagating larger fragments or full models. Our objective is to develop a principled understanding of how fragment size and dissemination patterns interact with the proportion of malicious nodes, and to design optimal fragmentation strategies that enhance robustness against adversarial behaviour.
To contribute effectively to this project, we highly value:
- Strong ML fundamentals and proficiency in ML implementation
[1] Noiseless Privacy-Preserving Decentralized Learning, PoPETS 2025
[2] Boosting Asynchronous Decentralized Learning with Model Fragmentation, WWW 2025
Free-rider detection in decentralized learning models using watermarks
Master’s thesis or MSc semester project
Contact: Jade Bourrée ([email protected]), Sayan Biswas ([email protected]), or Martijn de Vos ([email protected])
Decentralized learning has emerged as a promising alternative to centralized and federated paradigms for training machine learning models without relying on a central server. In fully decentralized settings, nodes collaboratively optimize a shared objective through local computations and peer-to-peer communication.
However, free-rider behavior remains a critical challenge: some nodes may benefit from the global model while contributing low-quality updates, random gradients, stale parameters, or no meaningful computation at all. Such behavior can degrade convergence, compromise fairness, and undermine trust in the system.
This project aims to design a robust watermark-based accountability framework to detect, quantify, and mitigate free-riding in decentralized learning systems. The core idea is to leverage model watermarking techniques to embed verifiable signals into the training process, enabling each client to claim legitimate safeguarding of intellectual property rights of the FL models.Research questions:– Can watermarking techniques developed in federated learning (e.g., [1]) be adapted to fully decentralized settings without a central coordinator?– How robust are watermark-based detection mechanisms against adversarial behaviors, such as collusion, gradient manipulation, or attempts to forge the watermark?
To contribute effectively to this project, we highly value:
- Strong ML fundamentals and proficiency in ML implementation
- Strong mathematical foundation and interest in probability theory, algebra, and analysis
[1] Li, Li, Xinpeng Zhang, Hanzhou Wu, Guorui Feng, and Weiming Zhang. “FareMark: Model-Watermark-Driven Free-Rider Detection in Federated Learning Model.” IEEE Internet of Things Journal (2025)