whoami

Ilyes Azouani

I study how deep models break and how to make them unbreakable — adversarial robustness, model security, and the geometry of learned representations applied to cyber threats.

quick description

I am a PhD candidate in machine learning and cybersecurity, in the mean time I am also the laboratory director of the AI Testing Laboratory of CLR Labs where I work with a beautiful team of passionate and brilliant people on evaluating neural networks from robustness to system’s cybersecurity and explainability, I am based in the Marseille area, between sea, mountains and buildings.

I am a proud husband, father of one and son.

This webpage is my $2^{nd}$ brain where you’ll find my articles, side projects, ideas…

If you want to know more please reach out via e-mail, I don’t have socials, might create something soon… and if you’re a student feel free to send a CV if you’re near Marseille and in the field of mathematics, deep learning and cybersecurity, if there’s an empty sit, you might find the right chair !

You might also find me under the pseudonym tonymagpie or tmgp. An ode to THPS II :wink.

github mail-me kaggle deep-ml rootme

medias

I like to make manimations, design, and other visually intriguing materials, I share them in this section also with significant events I’d appreciate to share with you.

manimations

design

scientific achievements

2026

Spherical caps and feature cones.

As part of my work in master’s degree I had the opportunity of publishing my first scientific work to the scientific community @SPRA26 The 6th Symposium on Pattern Recognition and Application. I published and presented an original view on why deep metric system (i.e. biometric verification) are vulnerable to adversarial examples and impostors in general via a geometrical modelisation of the problem of biometric verification. Adossed to Gilmer’s adversarial spheres we demonstrated how biometric verification is a problem of spherical caps, and overlapping identities on the surface of an hypersphere. This work led to a publication in SPIE Conference Proceeding.

Sadly I don’t have a recording of the presentation…

2025

Scientific communication at AMUSEC.

As part of my first year of master’s degree in applied mathematics I ought to give a talk on the threat of model inversion in learnt classifiers @AMUSEC2025 in beautiful Luminy,Marseille specifically at CIRM(International Center of Mathematics Encounters).

tmgp lab