Serge Durand

Serge Durand

PhD student in Trustworthy AI

CEA List LSL

INRIA Antique (ENS Paris)

About me

I am finishing a PhD on trustworthy AI with a defense in December 2025. I researched the use of over-approximations to help train, verify, and explain neural networks. I am broadly interested in making deep learning models more trustworthy.

On my free time I enjoy cooking (and eating), coffee (I tend to use freshly ground light/medium roast beans brewed with Aeropress), climbing (almost exclusively indoor bouldering), rugby (watching).

Interests

  • Artificial Intelligence
  • Neural Networks Robustness & Verification
  • Explainable AI

Education

  • Doctorat en Informatique (PhD), 2025

    ENS Paris - Saclay / CEA List

  • Master Parisien de recherche en Informatique - M2, 2021

    ENS Paris - Saclay / Paris 7 Diderot - now Université de Paris

  • Master Parisien de recherche en Informatique - M1, 2020

    ENS Paris - Saclay / Paris 7 Diderot - now Université de Paris

  • Licence de Mathématiques, mineure informatique, 2019

    Université Pierre et Marie Curie (UPMC) - now Sorbonne Université

Experience

 
 
 
 
 

PhD Student

CEA List, Laboratoire de Sécurité et Sûreté du Logiciel (LSL)

Nov 2021 – Dec 2025 Paris
Over-Approximating Neural Network for Verification, Robustness, and Explainability. PhD directed by François Terrier, co-supervised by Zakaria Chihani and Caterina Urban.
 
 
 
 
 

Intern

CEA List, Laboratoire de Sécurité et Sûreté du Logiciel (LSL)

Mar 2021 – Sep 2021 Paris
Neural Network Verification with PyRAT
 
 
 
 
 

Intern

Antique (INRIA / ENS Ulm lab)

Jun 2020 – Aug 2020 Paris
Verification of ACAS Xu networks using abstract interpretation

Recent Publications

Over-Approximating Neural Networks for Verification, Robustness and Explainability

We show in this thesis several applications of over-approximations of neural networks to improve trust in AI systems in three areas: verifications, robust training and (formal) explainability.
Serge Durand
Manuscript Slides PDF Slides Keynote Slides PPTX

On Using Certified Training towards Empirical Robustness

Combining certified and adversarial losses can help empirical robustness to local perturbations.
Alessandro De Palma, Serge Durand, Zakaria Chihani, François Terrier, Caterina Urban
PDF Code OpenReview

ReCIPH: Relational Coefficients for Input Partitioning Heuristic

An input partitioning heuristic for efficient neural network verification of low-dimensional input models.
Serge Durand, Augustin Lemesle, Zakaria Chihani, Caterina Urban, François Terrier
PDF Source Document