DataTerminology.comNEW
Also known as: Fast Gradient Sign Method
An attack that perturbs inputs in the sign of the loss gradient to create adversarial examples efficiently.
DESCRIPTION
The Fast Gradient Sign Method generates adversarial examples by taking a single step in the direction of the sign of the gradient of the loss with respect to the input. It is computationally efficient and serves as a baseline for robustness evaluation and adversarial training. FGSM typically targets L-infinity bounded perturbations with a chosen epsilon. Variants like BIM and PGD extend FGSM to iterative procedures for stronger attacks.
ALTERNATIVE DEFINITIONS
Goodfellow, Shlens, Szegedy
The fast gradient sign method (FGSM) uses the sign of the gradient of the loss function with respect to the input to generate adversarial examples.
Original definition linking attack direction to gradient sign.
Google Developers
FGSM adds a small perturbation in the direction of the sign of the gradient of the loss with respect to the input.
Plain-language explanation and practical framing for ML engineers.
RELATED TERMS
Adversarial Examples Adversarial Training Model Robustness Loss Gradient L-infinity Norm
COMPETING TERMS
PGD BIM
KNOWLEDGE AREAS
AI Data Science Testing & Validation
TARGET AUDIENCE
AI Engineer Data Scientist MLOps Engineer
EXTERNAL SOURCES
- Wikipedia - Adversarial machine learning
- arXiv - Original FGSM paper
- Google Developers - Adversarial examples overview
- PyTorch - FGSM tutorial with code
TOP QUESTIONS
What is FGSM?
FGSM is a single-step adversarial attack that perturbs inputs along the sign of the loss gradient to fool models.
Category: Definition
How do you control FGSM strength?
The epsilon parameter bounds the L-infinity norm of the perturbation. Larger epsilon yields stronger but more detectable attacks.
Category: HowTo
How does FGSM compare to PGD?
FGSM is faster but weaker, while PGD performs multiple iterative steps for stronger attacks at higher cost.
Category: Comparison
Is FGSM useful for training?
Yes, it can be used for adversarial training, though PGD-based training often provides stronger robustness.
Category: Application
What norms does FGSM target?
FGSM typically targets L-infinity constraints, but variants can be adapted to other norms.
Category: Concepts
What is a common mistake using FGSM?
Using a fixed epsilon without considering input scaling or perceptual visibility can produce unrealistic perturbations.
Category: Mistakes
Can FGSM be detected easily?
Large epsilon attacks are often detectable. Small, well-chosen perturbations can remain subtle.
Category: Risks
Where should I start learning FGSM?
Read the original paper and try official PyTorch or TensorFlow tutorials to understand gradients and constraints.
Category: HowTo
What is a hot topic around FGSM?
Developing adaptive evaluation protocols and comparing single-step vs. multi-step robustness remains active.
Category: Topic