Offre en lien avec l’Action/le Réseau : – — –/– — –

Laboratoire/Entreprise : Inria Saclay
Durée : 3 ans
Contact : cedric.eichler@insa-cvl.fr
Date limite de publication : 2024-09-01

Contexte :
This PhD thesis project is part of the French Priority Research Program and Equipment
(PEPR) on Cybersecurity, interdisciplinary Project on Privacy (iPoP) project involving several French
research teams working on data protection, from Inria, universities, engineering schools and the CNIL
(French National Commission on Information Technology and Civil Liberties). The PhD is proposed by
Petrus project-team at Inria Saclay and the PETSCRAFT project-team joint between Inria Saclay and
INSA CVL, which tightly collaborate in this large initiative on modeling privacy protection concepts and
on the design and deployment of explicable and efficient Privacy-Enhancing Technologies (PETs).

Sujet :
Objectives of the thesis.
The advanced inference capabilities of Large Language Models (LLMs) pose a significant threat to the privacy of individuals by enabling third parties to accurately infer certain personal attributes from their writings [1, 2]. Paradoxically, LLMs can also be used to protect individuals by helping them to modify their textual output from certain unwanted inferences [3, 4], opening the way to new tools. The ultimate objective of this thesis is to work towards an interactive chatbot-like tool for the sanitisation of text, to address applications including two which are especially investigated by our team: production of testimonies in the context of school bullying and work harassment, and participants feedback in participatory platforms. Through a preliminary investigation, we identified guidelines and main difficulties the successful PhD candidate will have to address for the sound development of such a
tool:
•A realistic adversary should be used to assess (residual) privacy risks. This poses two main challenges. Firstly, a realistic attacker cannot be generic but must take into account the vast auxiliary knowledge an attacker may possess (e.g. through fine-tuning or with the help of a dedicated ontology). Secondly, LLMs tend to always propose a guess which could be as likely as a random guess. Therefore, there is a need for a mechanism to estimate the likelihood of inferences.
•Designing and implementing a metric assessing the utility of a text (or the loss of utility due to sanitisation) is no trivial task. Design-wise, a proper metric should evaluate the amount of information
conveyed by a text relevant to its purpose (e.g. wrt testimonies, whether the victim/perpetrator are identifiable, etc). With regard to implementation, the assessment must be done automatically
without human intervention (e.g. through a LLM).
•Finally, an LLM-based sanitisation process must be proposed, limiting the capacity of the attacker to make inferences while maintaining the utility of the text. In a chatbot-like application, this
process can be iterative and interactive.

Initial roadmap.
The PhD project will start by the installation of open source LLMs such as Mistral or Arctic, and the implementation of the guidelines above, before focusing on the specialisation of the anonymisation solution to adapt it to different use cases and datasets.

Potential use-cases.
We will focus on two use cases: (1) the anonymous declaration or anonymisation of certain concepts in the context of school, university and work in general. This first use cases will be built with Inria’s partners in the context of the services responsible for investigating harassment cases that deal with anonymous witness statements and/or in the context of the labour market and job searches. (2) a second use-case is user feedback in participative platforms aimed at wellbeing, nutrition and health. This use case is still emerging and will be detailed during the PhD project.

References
[1] Kandpal, N., Pillutla, K., Oprea, A., Kairouz, P., Choquette-Choo, C., Xu, Z.: User inference attacks
on llms. In: Socially Responsible Language Modelling Research (2023)
[2] Staab, R., Vero, M., Balunovi ́c, M., Vechev, M.: Beyond memorization: Violating privacy via inference
with large language models. arXiv preprint arXiv:2310.07298 (2023)
[3] Staab, R., Vero, M., Balunovi ́c, M., Vechev, M.: Large language models are advanced anonymizers.
arXiv preprint arXiv:2402.13846 (2024)
[4] Tannier, X., Wajsb ̈urt, P., Calliger, A., Dura, B., Mouchet, A., Hilka, M., Bey, R.: Development and
validation of a natural language processing algorithm to pseudonymize documents in the context of
a clinical data warehouse. Methods of Information in Medicine (2024)

Profil du candidat :
Candidates must hold a master (or equivalent) in Computer Science.
The following skills are appreciated:
•Basic knowledge in LLMs/ML.
•Basic knowledge in privacy & anonimization.
•Proficiency in programming.

Candidates must be proficient in either french or english.

Formation et compétences requises :
Candidates must hold a master (or equivalent) in Computer Science.
The following skills are appreciated:
•Basic knowledge in LLMs/ML.
•Basic knowledge in privacy & anonimization.
•Proficiency in programming.

Candidates must be proficient in either french or english.

Adresse d’emploi :
Inria Saclay centre at Universit ́e Paris-Saclay, Turing building in Palaiseau (near Paris), France

Document attaché : 202406060959_PhD_position_LLM_Privacy.pdf

Offre en lien avec l’Action/le Réseau : – — –/– — –

Laboratoire/Entreprise : LITIS – Rouen
Durée : 3 ans
Contact : benoit.gauzere@insa-rouen.fr
Date limite de publication : 2024-10-01

Contexte :
Unlocking the Mysteries of the Brain with Graph-Based Machine Learning

In the fascinating world of neuroscience, understanding the brain’s intricate structure is key to unlocking the secrets of psychiatric and neurological disorders. Imagine if we could map the brain’s folds and curves to reveal patterns that indicate health or disease. This is precisely what our cutting-edge PhD project aims to achieve, leveraging the power of graph-based machine learning (GML) and Graph Neural Networks (GNN).

Project Overview

Our research focuses on enhancing the representation and analysis of neuroimaging data, particularly from MRI scans, using innovative GML techniques. By developing advanced models, we aim to identify individual traits such as gender and pathology with. What’s more, we are embedding principles of fairness into our models to ensure they are robust against variations in data acquisition and the natural diversity of brain structures.

Sujet :
Key Research Questions

– Hierarchical Information Analysis in Brain Graphs: How can we design GML models that effectively capture and utilize hierarchical information in brain graphs for better analysis of cortical folding patterns?
– Robustness to MRI Variations: Can a GML model trained on data from one MRI acquisition center generalize well to data from other centers, demonstrating robustness and enhancing the reproducibility of neuroimaging studies?
– Local Variation and Cognitive Functions: How can GML approaches help us identify and analyze local variations in brain anatomy, and what can these variations tell us about cognitive functions and neurological conditions?

Methodology and Resources

Our PhD candidate will have access to premier datasets, including:
– Human Connectome Project: Featuring top-quality MRI data from 1200 individuals.
– UK Biobank: Offering multimodal MRI data from over 10,000 individuals.

Supervisors and Research Environment

The project will be hosted at INSA Rouen, within the LITIS laboratory, and will be co-supervised by experts in the field:

Benoit Gaüzère
Guillaume Auzias
Sylvain Takerkart
Paul Honeine

Our multidisciplinary team brings together expertise in machine learning, computational anatomy, and neuroscience. The candidate will benefit from collaborations with leading research teams across multiple institutions.
How to Apply

Ready to embark on this exciting journey? Send your resume, academic results, and links to code or scientific papers to the following contacts. Please include “[FAMOUS]” in the subject line of your email:

Benoit Gaüzère: benoit.gauzere@insa-rouen.fr
Paul Honeine: paul.honeine@univ-rouen.fr
Guillaume Auzias: guillaume.auzias@univ-amu.fr
Sylvain Takerkart: sylvain.takerkart@univ-amu.fr

full offer is available here.

Profil du candidat :
Candidate Profile

We are seeking a passionate and dedicated PhD candidate with:

A Master’s degree in data science, computer engineering, or a related field.
Proficiency in Python programming.
Strong knowledge or experience in machine learning and data science.
Experience with graph structures is a plus.
High motivation and an interest in neuroscience.
Excellent reading, writing, and communication skills in English.

Formation et compétences requises :
Candidate Profile

We are seeking a passionate and dedicated PhD candidate with:

A Master’s degree in data science, computer engineering, or a related field.
Proficiency in Python programming.
Strong knowledge or experience in machine learning and data science.
Experience with graph structures is a plus.
High motivation and an interest in neuroscience.
Excellent reading, writing, and communication skills in English.

Adresse d’emploi :
LITIS INSA Rouen Normandie

Document attaché : 202406051242_PhD_project_Gauzere_Auzias.pdf

Offre en lien avec l’Action/le Réseau : – — –/– — –

Laboratoire/Entreprise : INRIA – EVERGREEN
Durée : 3 ans
Contact : diego.marcos@inria.fr
Date limite de publication : 2024-06-30

Contexte :
This PhD offer is funded by the GEO-ReSeT ANR project, representing a collaboration between Inria (team EVERGREEN, Montpellier) and Université de Paris Cité (team LIPADE, Paris).

Leveraging the large amounts of available geo-spatial data from different sources, the GEO-ReSeT (Generalized Earth Observation with Remote Sensing and Text) project has the objective to learn a rich representation of any geo-spatial location and convey a semantic representation of the information, by improving on existing models and providing a better experience to the end users. By using location on the Earth’s surface as the common link between different modalities, a geo-spatial foundation model would be able to incorporate a variety of data sources, including remote sensing imagery, textual descriptions of places, and other generic features.

Such a foundation model has the potential to open a set of all new possibilities in terms of Earth observation applications, by allowing for few or zero-shot solutions to classical problems such as land-cover and land-use mapping, target detection, and visual question answering. It will also be useful for a wide range of applications with a geo-spatial component, including environmental monitoring, urban planning and agriculture.
By leveraging several data modalities, this foundation model could provide a comprehensive and accurate understanding of the Earth’s surface, enabling informed decisions and actions. This will be particularly valuable for new potential users in sectors such as journalism, social sciences or environmental monitoring, who may not have the resources or expertise to collect their own training datasets and develop their own methods, thus moving beyond open Earth observation data and democratizing the access to Earth observation information.

Sujet :
The work to be conducted during the proposed PhD thesis will contribute to the ambition of the GEO-ReSeT ANR project by linking textual descriptions of places (e.g., collected from heterogeneous online sources, such as news articles or search engine results), to their approximate geo-location, a task known as geoparsing.

This text-location link will then be used in combination with other geospatial data modalities, with a focus on remote sensing data from sensors such as Sentinel-1 and -2, in order to train multi-modal models that are aware about the way in which people describe locations.

This will be done by first combining information stemming from different databases containing geographic named entities, such as Open Street Map, Wikipedia and gazetteers, such that geographic points or polygons can be linked to each named entity.

In a second step, a Natural Language Processing (NLP) pipeline will be developed to obtain the most likely geographic named entities that are referred to in any piece of text that describes a place.

With respect to existing Named Entity Recognition (NER) methodologies, in order to avoid restricting us to cases where entities’ names appear exactly as in the databases or gazetteers, we will leverage pre-trained Large Language Models (LLM) to resolve ambiguities and gather evidence towards the most likely entities that are being described in the text. Such an approach will be trained and validated by using the cases that do match the names in the gazetteer.

We will then move on, in collaboration with the rest of the GEO-ReSeT consortium, to train a multi-modal large language model (MMLLM) that will serve as a foundation model for Earth observation tasks.

This model will finally be evaluated on several agro-environmental tasks.

Application must be sent through the following link : https://recrutement.inria.fr/public/classic/en/offres/2024-07756

Profil du candidat :
Main activities
Description of the state-of-the-art in unstructured text geoparsing, with a focus on approaches leveraging LLMs.
Collection of a database of geographic named entities linked to their geographic footprint (e.g. point or polygon). Collection of a database of unstructured online text that is likely to contain a reference to a geographic location.
Development of an NLP pipeline to link each piece of geographic text to its likely geographic footprint.
Participate in the design and training of a multi-modal large language model (MMLLM) using remote sensing and geoparsed text.
Evaluation of the final model on two of the following case studies at a national or continental scale: ecosystem type mapping, crop type mapping or land-use mapping.

Formation et compétences requises :
Skills
Python programming.
Deep Learning with Python (preferably with Pytorch).
Experience with NLP.
Experience with GIS would be a plus.

Adresse d’emploi :
Montpellier, France