Inputs specially crafted to fool a machine-learning model into producing an incorrect output, while being indistinguishable from normal inputs to a human.
Adversarial examples are input data to which small, often imperceptible, perturbations have been deliberately added to induce a classification or prediction error in a machine-learning model. The phenomenon was demonstrated in 2014 by Goodfellow et al. on vision neural networks: a correctly classified panda image can be pixel-perturbed to be classified as a gibbon with high confidence, without the perturbation being visible to the naked eye. Adversarial robustness is now a central security criterion for any model deployed in adversarial environments: facial recognition systems, fraud detection, autonomous driving, insurance document processing. White-box attacks, where the attacker knows the model architecture and weights, are distinguished from black-box attacks, where only outputs are available. In the LLM context, adversarial examples take the form of carefully crafted prompts designed to bypass guardrails. For an insurer using a credit scoring or fraud detection model, vulnerability to adversarial examples constitutes a model risk to be integrated into the ORSA and algorithmic governance.
An auto insurer deploys a vision model to assess damage from claims photographs. A dishonest repair expert submits photos slightly altered by a digital filter that artificially reduces the damage estimate by 40%. The model is not robust to adversarial perturbations, leading to systematic underestimation of claims discovered several months later.
adversarial examples, exemples adversariaux, perturbations adversariales, adversarial inputs, exemples adverses