The SPRING lab offers project opportunities for BSc, MSc, and PhD students. We encourage interested students 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: 5th November 2025
Please, apply via Google form (login may be required). 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.
For all applicants, when filling the form, please use your institutional email so that we can get back to you.
For External students, i.e., students who are not from EPFL nor ETHZ, in addition to filling the application form, please also send an email to the supervisor(s) including (1) a self introduction within 5 sentences (2) how you are able to conduct this project as an external student (e.g., have you applied for an exchange/internship to EPFL or MPI-SP?) (3) apart from supervision, whether/how you need our support (e.g., internship salary, visa application, admin for exchange etc.)
We are gathering applications for projects during the Autumn Semester 2025 via the Google form.
Changes this year
Applications are processed in two rounds. For each round, we collect applications before the deadline. Then, we will get back to selected applicants during the corresponding “First Contact From Supervisors” period. If we do not get back to you during the indicated period, it means that we probably do not have space anymore.
We will make a mark on the project once it is taken. We strongly recommend that you apply as soon as possible for best consideration, since we expect most projects would be taken after the first round. However, we will leave the form open after the second round and consider all applications, if there are still available projects at that time.
For Master Projects (PDM)
For Research Projects
Important Notes:
Note that projects will be updated or added until 10th November 2025. We recommend that you check this page regularly for updates. You can modify your application at any point to select projects that you missed up until 5th December 2025.
If you encounter any technical issue, please get in touch with Saiid El Hajj Chehade.
Being anonymous when browsing a website is essential to preserve freedom of speech and information. To achieve this, one can use VPNs, or employ anonymous networks. However, the metadata of exchanged packets remains exposed. In website fingerprinting attacks [1], a passive attacker uses this metadata to predict, with machine learning techniques, which website a user has accessed. These attacks may allow governments or Internet Service Providers (ISPs) to monitor communications, thereby threatening user privacy.
Recent work in website traffic fingerprinting often relies on fixed monitored and unmonitored website lists drawn from popular ranking datasets. However, the composition of these lists can unintentionally introduce patterns that make classification artificially easier. For example, prior datasets such as BigEnough [2] rely on monitored websites drawn from highly ranked popular sites, while the unmonitored set contains thousands of unrelated websites with different temporal characteristics. This difference in timing behavior can leak information and give an unfair advantage to classifiers in open-world settings.
In this project, we aim to explore the following big questions:
The student will analyze existing datasets such as BigEnough and experimentally evaluate how different data selection strategies affect classifier performance in open and closed-world settings.
Requirements
Applying to this project
This PDM or 12 ECTS research project is aimed at one Master student. The student will work with Eric Jollès.
[1] Siby et al. “Evaluating practical QUIC website fingerprinting protections for the masses”
[2] Matthews et al. SoK: A Critical Evaluation of Efficient Website Fingerprinting Defenses
Applications such as maps, navigation and weather services rely on users sharing their location with a service provider in exchange for location-dependent information (weather at a location, navigation from one place to another). The goal of this project is to design a system for privacy-preserving location services and implement a prototype thereof. In this project, you will:
Requirements
Applying to this research project/master’s project (PDM)
This PDM or 12 ECTS research project is aimed at one Master student. The student will work with Christian Knabenhans
[1] Fung, E., Kellaris, G., & Papadias, D. (2015). Combining Differential Privacy and PIR for Efficient Strong Location Privacy. SSTD
[2] Yoo, J. S., Kim, T., & Yoon, J. W. (2025). Versatile and Fast Location-Based Private Information Retrieval with Fully Homomorphic Encryption over the Torus. CoRR, abs/2506.12761.
Remote therapies are useful to treat various psychological troubles. However, any data exchanged in such therapies, as well as the fact that such a therapy is even taking place, are very sensitive. In this project, you will explore and map the feasibility of using privacy-enhancing technologies for such settings.
Requirements
Applying to this research project/master’s project (PDM)
This PDM or 12 ECTS research project is aimed at one Master student. The student will work with Christian Knabenhans.
Since the very first paper in 1996, the research around censorship and its resistance in the computer security community is approaching the 30-year-old’s birthday [1]. In a recent work, we showed how a provably secure circumvention system proposed in the literature failed to achieve empirical security under a real-world threat model. This mismatch between theoretical guarantees and empirical security is, on the one hand, alarming when we consider the risk of censor retaliation on circumventing users; on the other hand, begging the question:
Maybe we aimed for a high score on the wrong exam, but what is the right, or rather, a good enough exam?
In this project, we will take a critical look at circumvention systems from the literature, focusing on whether the evaluation in the paper actually provides a fair examination of claimed properties. Some questions to start with can be: What is the threat model? Does the threat model make sense in any specific censorship context? Are we evaluating in a limited manner which does not reflect how a rational censor would behave in reality? What are the assumptions that must hold for provable properties? To what extent they build up system properties? How to step towards a critical, fair, systematic evaluation?
This project is research-oriented, in the sense that we aim for a publication based on the thesis. Hence, it is particularly suitable for students who would be interested in figuring out whether academia is the way to go, or building up research skills.
Requirements
Applying to this project
This PDM or 12 ECTS research project is aimed at one Master student. The student will work with Boya Wang.
[1] https://censorbib.nymity.ch/
The bulk of the web security and privacy literature focuses on measuring the number of tracker requests encountered by web users and using it as a metric for the privacy harms to users on these websites. However, web advertisement and tracking services (ATS) can share user information with each other – forming “coalitions” – which exponentially increases user browsing pattern leakage. Learning from such a graph would allow us to measure privacy leakage more accurately (e.g, belief propagation modeling [1]) and uncover patterns in websites that share the same ATS coalition.
Since ATS information sharing occurs between their backends, we cannot measure ATS-to-ATS links directly by monitoring browser network activity. Some studies use proxy metrics to uncover some connections between trackers in the browser: e.g, cookie syncing [2], redirect chains [3], etc. Other studies rely on information disclosed by ATSs on their home pages in their privacy policies [4] or by publishing their coalitions [5]. In this project, we aim to answer the following big questions:
Recently, Ghostery – a prominent ad-blocker company – open-sourced a dataset of hand-labeled tracker information: TrackerDB [6]. It contains information about the parent companies, the type of ATS, and the ad-blocker rules used to block them. By identifying these trackers on webpages and analyzing their publicly available information, we can connect different ATS members of TrackerDB and build TrackerGraph.
What can we learn from building TrackerGraph? What about new information-sharing techniques? What about unknown “coalitions” that could be malicious?
This project follows a clear research methodology and is expected to yield results that are publishable. This means this project is best suited for students who are eager to participate in all the steps of the research method (literature review, solution design, experimental setup, analysis, and (maybe) writing).
Requirements
Applying to this project
This PDM is aimed at one Master student. The student will work with Saiid El Hajj Chehade
[1] https://en.wikipedia.org/wiki/Belief_propagation
[2] Panagiotis Papadopoulos, Nicolas Kourtellis, and Evangelos Markatos. 2019. Cookie Synchronization: Everything You Always Wanted to Know But Were Afraid to Ask. In The World Wide Web Conference (WWW ‘19).
[3] Iqbal, Umar, et al. “Khaleesi: Breaker of advertising and tracking request chains.” 31st USENIX Security Symposium (USENIX Security 22). 2022.
[4] Brookman, Justin, et al. “Cross-device tracking: Measurement and disclosures.” Proceedings on Privacy Enhancing Technologies (2017).
[5] https://id5.io/
[6] https://github.com/ghostery/trackerdb/
The advertisement and tracking ecosystem (ATSs) classifies users into two categories: low-value users (LVUs) and high-value users (HVU), to determine where to allocate their advertising budgets [1]: LVUs are users that are less likely to be receptive to ATS campaigns (e.g, bots, fresh browsers, etc.), whereas HVUs represent the average web consumer who spends a longer time on the web, allowing ATSs to draw a better picture about their preferences and serve them targeted advertisements.
Prior web measurement studies suggest that ATSs exhibit different behaviors on webpages when visited by LVUs (specifically, bots and research-instrumented browsers) compared to HVUs (real people with rich browser usage) [2]. As such, automated web measurements are not necessarily representative of what people observe in the wild, and can be blind to more advanced forms or volumes of tracking. Additionally, using fresh browser sessions in automated measurements prevents us from determining how trackers react differently to sub-communities of users (based on gender, consumer-focused vs. information-focused, career, age, etc.).
Can we use LLMs to generate subsets of user communities and simulate their browsing patterns?
In this project, we will (1) discover the key axes of a user profile that trackers are most reactive to; (2) explore sustainable methods to generate these user profiles by asking the LLM to generate simulated search queries that the user they represent might ask; and lastly (3) evaluate the “quality” and “representativeness” of the profiles.
Requirements
Applying to this project
This PDM or 12 ECTS research project is aimed at one Master student. The student will work with Saiid El Hajj Chehade
[1] https://patents.google.com/patent/US12260451B2/en
[2] https://dl.acm.org/doi/abs/10.1145/3366423.3380104