Published October 1, 2025
Library
graduated
Robust-AI
PythonPackage
Maintainer:IRT-SystemX
Description
Train neural network models that are robust to adversarial attacks
Implementation of the Projected Gradient Descent (PGD) attack as the primary technique to estimate and assess the robustness of a model. The models are built upon an industrial use-case (Not-limited to) of visual inspection and tested with the integration of Cracked pavement detection (SDNET2018 dataset). We implement several state-of-the-art adversarial training methods on different DNNs architectures to enhance the model's performance against adversarial attacks.
Owner:IRT-SystemX
Keywords:ML-trainingadversarial-attackrobust-ai
CONTEXT
Robustness is a key point to build Trustworthy AI systems. It ensures that
the behaviour of an AI component in production is stable when facing to
different kind of perturbations. Adversarial attacks are a kind of intentional
pertubation created by malicious people making subtle modifications to AI
input data to deceive an artificial intelligence system, while remaining
imperceptible to human observers.
VALUE PROPOSITION
Robust-ai is a library of training methods designed to improve the robustness of computer vision models. It proposes some architectures and many training methods to easily train an AI model that is robust to adversarial attacks.
WHEN TO USE IT
It shall be used by data scientist for model development during the activity 'Develop and acquire ML models'.
RESOURCES
