FGM

Gold definitionUpdated Apr 2, 2026

Definition

FGM (Fast Gradient Method) is a prominent adversarial attack technique that generates small, input-dependent perturbations to mislead machine learning models. It works by adding a perturbation proportional to the sign of the loss gradient with respect to the input, aiming to maximize the model's prediction error.

At a glance

Executive summary

FGM is a method to create subtle, targeted changes to data inputs that trick AI models into making mistakes. It's used to test how robust AI systems are, revealing that even advanced models can fail significantly when faced with these 'adversarial' inputs, despite performing well normally.

TL;DR

FGM is a technique that makes tiny, specific changes to data to fool AI models, helping researchers understand how easily these models can be tricked.

Key points

Generates adversarial examples by adding a perturbation based on the sign of the loss gradient with respect to the input.
Solves the problem of evaluating and understanding the robustness vulnerabilities of deep learning models.
Used by researchers and ML engineers in adversarial machine learning, security, and model robustness studies.
Unlike random noise, FGM perturbations are specifically crafted to maximize model error, making them more effective than equal-norm random shifts.
Research trend focuses on developing more robust models and effective adversarial training methods to defend against FGM and similar attacks.

Use cases

Evaluating the robustness of image classifiers against imperceptible changes that cause misclassification.
Testing the security of autonomous driving systems by generating adversarial road signs or objects.
Benchmarking the resilience of natural language processing models to subtle text alterations.
Assessing the vulnerability of AI-based error correction decoders in communication systems to channel noise shifts.
Developing and validating adversarial training techniques to improve model defenses against gradient-based attacks.

Also known as

Fast Gradient Sign Method, FGSM

FGM