Introduction to Robust Machine Learning with Geometric Methods for Defense Applications

Pierre-Yves Lagrave (1) , Frédéric Barbaresco (2)

1 Thales Research and Technology [Palaiseau]
2 Thales LAS France

This paper aims at motivating the use of geometrically informed Machine Learning algorithms for Defense applications by providing intuitions with respect to their underlying mechanisms and by shedding light on successful applications such as remote sensing imagery, radar Doppler signal processing, trajectory prediction, physical model simulation and kinematics recognition. We in particular discuss the use Equivariant Neural Networks (ENN) which achieve geometrical robustness by-design and which also appear more robust to adversarial attacks. We will also highlight how Lie Group Statistics and Machine Learning techniques can be used to process data in their native geometry. Both technologies have a wide range of applications for the Defense industry and we generally believe that exploiting the data geometry and the underlying symmetries is key to the design of efficient, reliable and robust AI-based Defense systems.

Mots clés: Equivariant Neural Networks; Geometric Deep Learning; Lie Group Statistics and Machine Learning; Robustness-by-design

Domaines: Intelligence artificielle [cs.AI]; Apprentissage [cs.LG]; Théorie des groupes [math.GR]; Machine Learning [stat.ML]

Pierre-Yves Lagrave, Frédéric Barbaresco. Introduction to Robust Machine Learning with Geometric Methods for Defense Applications. 2021. ⟨hal-03309807⟩