Multi-Agent Planning under Knowledge and Communication Constraints
Supervisor: Rayna Dimitrova
Project Identifier: RD_MultiAgent
Description: Coordinating multiple intelligent agents in dynamic and uncertain environments poses fundamental challenges. While centralized approaches assume full observability, knowledge of other agents’ strategies, and unrestricted communication, real-world multi-agent systems, such as autonomous vehicles, distributed robots, or sensor networks, must operate with limited knowledge and constrained communication bandwidth.
In this project, you will investigate the automatic construction of coordination strategies for multi-agent systems under knowledge and communication constraints. The research will explore how epistemic and information-theoretic principles—including mutual information, entropy, communication cost, and knowledge representation—can be used to formalize and optimize the trade-off between knowledge sharing and communication efficiency.
Depending on the internship’s focus, the work may involve developing or adapting algorithms within frameworks such as (partially observable) stochastic games or epistemic reasoning. The expected outcomes include a characterization of communication–efficiency trade-offs and novel methods for decentralized coordination inspired by epistemic and information-theoretic insights.
References: "Synthesis of Communication Policies for Multi-Agent Systems Robust to Communication Restrictions". Saleh Soudijani, Rayna Dimitrova. IJCAI 2025 (https://www.ijcai.org/proceedings/2025/0030.pdf)
Internship Location: St. Ingbert
Prerequisite: none
Online Behavioral Advertising on Social Media
Supervisor: Jane Im
Project Identifier: JI_OBA
Description: In the course of this internship, you will be conducting interviews to understand how people perceive social media companies’ business models, especially how they profit from online behavioral advertising. The focus will be on business models and revenues, rather than how OBA works.
Internship Location: St. Ingbert
Prerequisite: Ideally, you have basic experience with semi-structured interviews
Fairness and Consent
Supervisor: Jane Im
Project Identifier: JI_Fairness
Description: In this internship. you will be conducting experiments to understand how people tradeoff 1) fairness-related benefits to a whole organization or society and 2) benefits to themselves when making consent decisions in regards to being included in AI training datasets. For example, very privacy-aware people will definitely try to opt out of being included in datasets, but this means they could be less represented. How would different kinds of framing around consenting impact people’s consent decisions? The student will first conduct a literature review of relevant papers and come up with different ways to frame the implications of consenting to being included in model training datasets with respect to fairness, privacy, etc. Then, they will design and conduct an online experiment.
Internship Location: St. Ingbert
Prerequisite: You should have completed a statistics course and basic knowledge in programming with R.
Online Behavioral Advertising on Social Media
Supervisor: Jane Im
Project Identifier: JI_Business_models
Description: In the course of this internship, you will be conducting interviews to understand how people perceive social media companies’ business models, especially how they profit from online behavioral advertising, as well as alternative business models (e.g., subscriptions, donation-based, paying for privacy features). The focus will be on understanding why people perceive tech companies’ business models and revenues in a certain way---for example, it’s likely that many people don’t want to pay for using social media. But why is that the case, when many users say they do not want to see ads on platforms?
Internship Location: St. Ingbert
Prerequisites: Ideally, the candidate has basic experience with semi-structured interviews.
Relational Machine Learning
Supervisor: Rebekka Burkholz
Project Identifier: RB_ML
Description: Modern AI systems increasingly rely on multiple interacting agents — from language model ensembles to autonomous decision-making teams. Yet, most multi-agent systems overlook the potential benefits of diversity, specialization, and trust. This project explores how cascade dynamics from opinion formation models can be leveraged to improve coordination and collaboration among AI agents.
You will work on designing cascades of communication and influence between heterogeneous agents that differ in competency, specialization, and reliability. Key research directions include:
- How does agent diversity affect collective decision making?
- Can trust and reputation mechanisms stabilize collaboration?
- What emergent behaviors arise from cascades of opinion updates among specialized agents?
- Can agents learn to trust each other and collaborate effectively?
The goal is to uncover principles of collective intelligence in agentic AI, moving beyond traditional ensembling toward adaptive, trust-aware cooperation among autonomous models. We will show that indeed the whole is greater than the sum of its parts.
Internship Location: St. Ingbert
Prerequisite: Basic knowledge of Deep Learning; Pytorch; experience with LLMs or GNNs is a plus.
Mind the Bits: Optimizers that Survive Quantization and Compression
Supervisor: Rebekka Burkholz
Project Identifier: RB_ML2
Description: Modern AI relies heavily on large models that consume enormous amounts of compute and memory. Compression and parameter quantization are common approaches to reduce the model size of trained models but they often degrade model performance severely. The main question is whether we can find better model representations that are more robust to compression and quantization. This requires us to think about efficiency from the start and thus during training.
This project investigates how optimizer dynamics influence quantization and compression sensitivity and explores strategies to develop quantization/compression-aware optimizers that inherently produce more resilient models. Key research questions include:
- What functional or geometric properties of learned representations make them easier to quantize or compress?
- How do optimizer choices affect quantization/compression error propagation?
- Can we introduce regularization or gradient-scaling techniques to suppress outliers (that are difficult to handle during compression/quantization) during training?
In collaboration, you will combine empirical experiments on standard benchmarks with theoretical insights into optimization dynamics, contributing to the broader goal of training models that are both accurate and hardware-efficient by design.
Internship Location: St. Ingbert
Prerequisites: Basic knowledge of deep learning; Pytorch.
Privacy-Preserving Authenticatio
Supervisor: Wouter Leuks
Project Identifier: WL_Privacy_Systems
Description: In this project you will design a new privacy-preserving system for a real-world application. In the past we have worked on projects with humanitarian organizations, public health, and investigative journalists. A key challenge in such systems is that sometimes information must be hidden even from people that have access to our mobile devices.
In this project, you will design a secure authentication method that can hold up to powerful adversaries. This project will require thinking about what an adversary can do (e.g., read the full mobile device? Interact with it?) and how to design a protection mechanism that is easy to use (e.g., maybe remembering a 128-bits string is not very useable.) A good solution will require some creative use of cryptography, while always keeping usability in mind.
We are also open to other ideas for privacy-preserving systems, and would be happy to discuss these with you.
Internship Location: St. Ingbert
Prerequisite: This project requires a creative mind, the ability to think about systems as a whole, and a reasonable foundation in cryptography and systems.Low-Cost Anonymous Communications
Supervisor: Wouter Leuks
Project Identifier: WL_Privacy_Comms
Description: The ability to transfer data anonymously over a network connection (e.g., via Tor) is very often essential to maintain the anonymity of a system as a whole. Tor, and other existing designs, however, are hard to integrate into applications, and come with considerable overhead.
In this project, you will build and evaluate low-cost anonymity systems that are (1) easy to integrate; (2) offer reasonable protections (better than VPNs); (3) have low operational cost. An ideal candidate for this project will be able to propose new ways to maintain anonymity (e.g., by combining existing ideas in novel ways) and is able to implement and evaluate their designs to show practicality.
Internship Location: St. Ingbert
Prerequisite: This project requires a strong understanding of network and systems. Knowledge of anonymous communication systems is a plus, but not required. It might help to know some basic cryptography as well.
Evaluating general-purpose functionality of synthetic data
Supervisor: Ana-Maria Cretu
Project Identifier: AMC_Synthetic1
Description: Governments and businesses are seeking solutions to provide access to private datasets in order to support innovation while protecting the privacy of individuals, e.g., in the healthcare and financial domains. Synthetic data, consisting of artificial records with similar statistical properties as the real data, has emerged as a promising approach to share data while preventing the re-identification of individuals in the data [1]. It is often claimed that synthetic data enables general-purpose downstream uses of the data, just like the raw data, while preserving privacy.
The goal of the internship is to evaluate whether synthetic data provides a better privacy-utility trade-off than tailored data releases. The student will select a dataset, some relevant use cases for the dataset, and a synthetic data generator and compute the privacy-utility trade-off of this generator. Then, the student will compare this trade-off with the trade-off achieved by data releases tailored to the use cases, e.g., in the case of classification as the use case, a tailored data release is a machine learning classifier. The project will involve (1) evaluating privacy using state-of-the-art attacks [2], (2) implementing or re-using differentially private algorithms for each data release [3], and (3) computing privacy-utility trade-off curves for every use case.
References:
- Gadotti, A., Rocher, L., Houssiau, F., Creţu, A. M., & De Montjoye, Y. A. (2024). Anonymization: The imperfect science of using data while preserving privacy. Science Advances, 10(29), eadn7053.
- Carlini, N., Chien, S., Nasr, M., Song, S., Terzis, A., & Tramer, F. (2022, May). Membership inference attacks from first principles. In 2022 IEEE Symposium on Security and Privacy (SP) (pp. 1897–1914). IEEE.
- Yousefpour, A., Shilov, I., Sablayrolles, A., Testuggine, D., Prasad, K., Malek, M., ... & Mironov, I. (2021). Opacus: User-friendly differential privacy library in PyTorch. arXiv preprint arXiv:2109.12298.
Internship Location: St. Ingbert
Prerequisite: The student should be comfortable coding with Python. Ideally, the student should be familiar with security and privacy notions from having completed a course on computer security, network security, or privacy, and with machine learning notions from having completed
a course on machine learning, statistics, or probabilities.
Efficient privacy auditing of synthetic data
Supervisor: Ana-Maria Cretu
Project Identifier: AMC_Synthetic2
Description: Governments and businesses are seeking solutions to provide access to private datasets in order to support innovation while protecting the privacy of individuals, e.g., in the healthcare and financial domains. Synthetic data, consisting of artificial records with similar statistical properties as the real data, has emerged as a promising approach to share data while preventing the re-identification of individuals in the data [1]. However, synthetic data does not solve all privacy issues, as it does not necessarily protect against inference attacks [2]. In an inference attack, an adversary combines information they have about an individual with the synthetic data to infer new information, such as their membership in a sensitive dataset or a sensitive attribute [2]. Practitioners use inference attacks to audit the vulnerability of synthetic data before releasing it. However, existing inference attacks are computationally very expensive to run [3]. The goal of the internship is to develop a more efficient inference attack. The student will implement a new inference attack, evaluate it against previous approaches, and attempt to improve its efficacy and efficiency.
References:
- Gadotti, A., Rocher, L., Houssiau, F., Creţu, A. M., & De Montjoye, Y. A. (2024). Anonymization: The imperfect science of using data while preserving privacy. Science Advances, 10(29), eadn7053.
- Stadler, T., Oprisanu, B., & Troncoso, C. (2022). Synthetic data–anonymisation groundhog day. In 31st USENIX Security Symposium (USENIX Security 22) (pp. 1451–1468).
- Meeus, M., Guepin, F., Creţu, A. M., & De Montjoye, Y. A. (2023, September). Achilles’ heels: Vulnerable record identification in synthetic data publishing. In European Symposium on Research in Computer Security (pp. 380–399). Cham: Springer Nature Switzerland.
Internship Location: St. Ingbert
Prerequisite: The student should be comfortable coding with Python. Ideally, the student should be familiar with security and privacy notions from having completed a course on computer security, network security, or privacy, and with machine learning notions from having completed
a course on machine learning, statistics, or probabilities.
Designing Metadata-Protected Communication Systems
Supervisor: Sajin Sasy
Project Identifier: SS_MPCS
Description: End-to-end (E2E) encryption enjoys widespread deployment today and serves to protect the contents of our communications online. However, systems that rely on just E2E still leak information (or metadata) about who is communicating with whom, how much, and how often. Such metadata often endangers journalists, whistleblowers, and marginalized communities of our society. Consequently, several designs have been proposed for Metadata-Protected Communication Systems (MPCS) over the last decade.In this internship, you will work on new ideas and designs to improve the current state of MPCS.
Internship Location: St. Ingbert
Prerequisites:
- Successfully completed introductory courses in or relating to i) design and analysis of algorithms, ii) cryptography, and iii) security.
- Strong mathematical background and eagerness to learn more.
- Background/Introductory reading: SoK: Metadata-Protected Communication Systems