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 Fall 2026.
Privacy Vulnerability Transfer in Model Merging
Master’s thesis or MSc semester project
Contact: Milos Vujasinovic ([email protected])
Model merging has emerged as a prominent technique for unlocking multitask capabilities in language and vision-language models at a fraction of the cost of full fine-tuning. It is also increasingly used to empower on-device models to deliver stronger performance under tight compute and memory budgets. Despite its rapid adoption, the privacy implications of merging remain poorly understood: when weights or task vectors from independently trained models are combined, it is not clear to what extent the privacy vulnerabilities of each constituent model transfer into, or are amplified within, the merged model.
This project investigates exactly that question. The student will study how privacy vulnerabilities propagate through model merging and task arithmetic [1], the foundation on which many modern merging techniques are built, and characterize the resulting risks under both traditional membership inference attacks [2] and language-model-specific membership inference and extraction attacks [3], with a particular emphasis on the latter. Concretely, the work will involve fine-tuning language models, constructing controlled merging setups, implementing and adapting state-of-the-art membership inference attacks against merged decoder-based language models, and analyzing how factors such as merging coefficients, the number of merged models, and task similarity influence vulnerability transfer.
RequirementsThis project is prerequisite-heavy. Applicants should have solid hands-on experience working with decoder-based language models (training, fine-tuning, and inference at minimum) and a general knowledge of privacy in machine learning, familiarity with membership inference and differential privacy (if we end up looking at it) is a plus. Comfort with PyTorch, the Hugging Face ecosystem is expected. Prior exposure to model merging and task vectors will be considered favorably.
References[1] Ilharco, G., Ribeiro, M. T., Wortsman, M., Gururangan, S., Schmidt, L., Hajishirzi, H., & Farhadi, A. (2022). Editing models with task arithmetic. arXiv preprint arXiv:2212.04089.[2] Yeom, S., Giacomelli, I., Fredrikson, M., & Jha, S. (2018). Privacy risk in machine learning: Analyzing the connection to overfitting. In 2018 IEEE 31st Computer Security Foundations Symposium (CSF) (pp. 268–282). IEEE.[3] Carlini, N., Tramèr, F., Wallace, E., Jagielski, M., Herbert-Voss, A., Lee, K., Roberts, A., Brown, T., Song, D., Erlingsson, Ú., & Oprea, A. (2021). Extracting training data from large language models. In 30th USENIX Security Symposium (USENIX Security 21) (pp. 2633–2650).