On this page you can find our offering for semester projects. Students who are interested in doing a project are encouraged to have a look at the Thesis & Project Guidelines from the MLO lab, where you will gain an understanding about what can be expected of us and what we expect from students.
Last Update: 18th May 2022
Please, apply via Google form. You will need to specify which project(s) you are interested in, why you are interested, and if you have any relevant experience in this area. To access the form you need to login with your EPFL email address.
We process students in two rounds. We collect applications for projects for the first round until 1st June 2022. We will get back to you about your application in the first week after the deadline. If we do not get back to you during the indicated period, this means that we unfortunately did not have space.
We will leave the form open for late applications. If by 11th July 2022 there are still projects open we will consider all applications submitted by then. We strongly recommend that you apply as soon as possible for best consideration. We expect many projects to be taken after the first round.
External students: Students who are not from EPFL should get in touch with the supervisors of the project(s) via email.
Early deadline: 1st June 2022
First Contact with Supervisors: 2nd June 2022 - 9th June 2022
Late deadline: 11th July 2022
First Contact with Supervisors: 12th July 2022 - 19th July 2022
If you encounter any technical issue please get in touch with Laurent Girod.
Substantial research effort has been devoted to investigating the adversarial capabilities of parameter servers in federated learning. However, currently proposed attacks are either applicable only on a limited set of unrealistic system configurations or trivial to detect by aware users. Ultimately, the ineffectiveness of these attacks on realistic and representative scenarios provides only a false sense of security to the community.
With this project, we aim to unveil the true capabilities of a malicious parameter server in federated learning, demonstrating the fundamental impact that it can have on users’ privacy. Here, we move beyond trivial attack strategies previously defined in the literature; furthermore, we expand our study to extensions of the federated learning protocol such as personalized federated learning and other emerging approaches.
Further details This semester project is aimed at one MSc student. The student will work with Dario Pasquini.
Tabular data is a rare domain in which Deep Learning has not been widely used yet. Traditionally, the performance of neural networks has proved to be less good than Decision Tree-based algorithms (gradient boosting, random forest, etc.). Recently, new approaches (e.g. TabNet) capable of the same or even better performance have appeared. Although these models show promising results, they also inherit a significant drawback of DNN – vulnerability to adversarial examples. Since image and text domains are better covered in the adversarial ML literature, specialised benchmarks for adversarial robustness exist for these domains (e. g. Robustbench. The goal of the project is developing adversarial robustness evaluation pipeline which is suitable for the tabular data.
Further details This semester project is aimed at one MSc student (BSc can be considered as well). The student will work with Klim Kireev.
Federated learning (FL) has gained significant attention for enabling collaborative machine learning among numerous data holders without sharing the data. Yet, it is shown that FL is not robust against several privacy attacks (e.g., membership inference) due to the shared intermediate values. We focus on several challenges of the full FL pipeline (before, during, or after training the FL model) and tackle the problem of privacy leakage during the pipeline.
In this project, our aim is to find a suitable algorithm for tackling the hyperparameter tuning challenge in FL while protecting the parties’ input data. We leverage several FL libraries that are already implemented. In agreement with the student, possible projects may involve one of the following:
Further details Project type: Semester project (8 or 12 credits). The student will work with Sinem Sav.
Collaborative learning allows distinct participants to train a joint model, without their local training data leaving the devices. Such approach has gained a lot of attention because of its privacy and security claims. See Google’s paper for an example. However, those claims rely on a “good” setup phase, which includes the choice of graph connecting users, initial parameters, etc.
In this project, the student is expected to challenge existing collaborative learning designs and their (sometimes not-so realistic) assumptions, and explore the privacy-utility trade-offs available during the setup phase.
This semester project is aimed at one MSc student (BSc can be considered as well). The student will work with Mathilde Raynal.
Note: Other labs with cool projects on Machine Learning (we can always consider co-supervision):
Homomorphic Encryption has recently become a trending cryptographic approach to enable privacy and utility in data processing pipelines. This kind of encryption enables one to execute computation on the ciphertext directly to obtain an encryption decrypting to the result as if it had been executed on the decrypted data.
Amongst the most trending homomorphic schemes are lattice-based constructions [BGV, BFV, CKKS] that can support both additions and multiplications. A new approach combines such schemes with multiparty homomorphic encryption to make it even more practical [MHE].
However, such data-analysis pipelines are in the honest-but curious model: the different parties might try to infer information but they will follow the protocol as specified. A new research direction is thus to explore how such pipelines could support more realistic threat models where parties might deviate from the protocol.
In this project, we will explore, implement, and evaluate potential approaches to port MHE pipelines to stronger threat models. Promising directions include proof systems and zero-knowledge proofs, actively secure multi-party computation, and more.
Project type: MSc Semester project (12 credits). The student will work with Sylvain Chatel.
[BFV] Fan et al., “Somewhat practical fully homomorphic encryption”, ArXiv 2012
[BGV] Brakerski et al., “(Leveled) fully homomorphic encryption without bootstrapping”, TOCT 2014
[CKKS] Cheon et al., “Homomorphic encryption for arithmetic of approximate numbers”, ASIACRYPT 2017
[MHE] Mouchet et al., “Multiparty homomorphic encryption from ring-learning-with-errors”, PETs 2021
The SPRING lab has been working with the International Committee of the Red Cross (ICRC) to design a privacy-friendly aid distribution system. Historically, humanitarian organisations have often provided fixed aid packages to beneficiaries (eg., three packets of rice). However, fixed packages make it impossible to adjust aid for household composition, nor do they enable recipients of aid to themselves select which products they like to receive.
To address this problem, humanitarian organisations need to switch to either cash distributions or wallets. Using a wallet, aid recipients can “pay” for the products that they want at different vendors. Earlier, we designed a privacy-friendly aid distribution system for fixed distributions. In this project, you will instead design a privacy-friendly wallet system that will support multiple vendors.
Applying to this project