Internship Research - Limiting Information Leakage In Llm-Based Systems For Network Applications H/F INRIA

Internship Research / Limiting Information Leakage in LLM-Based Systems for Network Applications
Le descriptif de l'offre ci-dessous est en Anglais
Type de contrat : Convention de stage

Niveau de diplôme exigé : Bac +2 ou équivalent

Fonction : Stagiaire de la recherche

A propos du centre ou de la direction fonctionnelle

Created in 2008, the Inria Saclay Center is located at the heart of the Paris-Saclay scientific and technological excellence cluster, which alone accounts for 15% of French research. Serving the development of the Université Paris-Saclay and the Institut Polytechnique de Paris, the Inria Saclay center employs 80 people in research support services and 500 scientists of 54 nationalities.

Benefiting from continuous growth, the center now has a total of 42 project-teams and two in the process of being created, including 21 jointly with the Institut Polytechnique de Paris, 16 with the Université Paris-Saclay, as well as 7 Inria EPs, including one in collaboration with Onera and one with the Pôle Universitaire Centre Val de Loire. These research teams are spread over more than ten sites.

Mission confiée

Context

Large language models are increasingly used to handle user requests that contain personal or sensitive information. While the model processes such a request, information from the input can be exposed beyond what is strictly needed to answer it. This raises concerns about privacy and data minimization.

Privacy in LLMs has been studied from several angles: memorization and extraction of training data [1], what it means for a language model to preserve privacy [2], contextual integrity in LLM behavior [3], and privacy-preserving reasoning [4].

Objectives

The internship will study how LLM-based systems can complete a user's request while limiting the exposure of sensitive information contained in it, with network applications as the target domain. The main questions are:

- How can the information flows of an LLM-based system be modeled?
- How can privacy leakage be formally defined and measured?
- What mechanisms reduce leakage while preserving the quality of the result?

Work plan
- Review the literature on privacy in LLMs and on privacy-preserving mechanisms, such as differential privacy [5].
- Define a threat model and a privacy-leakage metric, and select a representative network use case.
- Design and implement an approach to reduce leakage.
- Evaluate it experimentally against baselines, along a privacy-utility trade-off.

Candidate profile

- Interest in privacy and security, with knowledge of differential privacy as a plus
- Experience with LLMs and solid Python skills
- Knowledge of networking is a plus

References

[1] N. Carlini et al., Extracting Training Data from Large Language Models, USENIX Security 2021.
[2] H. Brown et al., What Does it Mean for a Language Model to Preserve Privacy?, ACM FAccT 2022.
[3] N. Mireshghallah et al., Can LLMs Keep a Secret? Testing Privacy Implications of Language Models via Contextual Integrity Theory, ICLR 2024.
[4] SALT: Steering Activations towards Leakage-free Thinking in Chain of Thought, arXiv:2511.07772, 2025.
[5] C. Dwork and A. Roth, The Algorithmic Foundations of Differential Privacy, Foundations and Trends in Theoretical Computer Science, 2014.

Avantages

- Subsidized meals
- Partial reimbursement of public transport costs
- Leave: 7 weeks of annual leave + 10 extra days off due to RTT (statutory reduction in working hours) + possibility of exceptional leave (sick children, moving home, etc.)
- Possibility of teleworking and flexible organization of working hours
- Professional equipment available (videoconferencing, loan of computer equipment, etc.)
- Social, cultural and sports events and activities
- Access to vocational training
- Social security coverage

Rémunération

Gratification