Adversarial Attack
A security attack technique that causes machine learning models to produce erroneous outputs by adding small perturbations to input data. The two principal categories are white-box attacks, in which the attacker has access to the model's internal structure, and black-box attacks, in which the attacker can only observe model outputs.
Timeline (last 90 days)
Counterexamples generated by standard adversarial attacks are independent and scattered, making it difficult to systematically cover the failure boundary; whereas the counterexample sets generated by the enumeration-guided method are more compact and representative, capturing the failure boundary more precisely.
The adversarial attack phenomenon was first discovered by Szegedy et al. in 2013, after which Goodfellow et al. proposed classic attack methods such as FGSM.
The core principle of adversarial example attacks is to apply imperceptibly small perturbations to input data that are nearly undetectable to the human eye, causing a trained deep learning model to produce completely incorrect outputs.
All Facts (3)
The core principle of adversarial example attacks is to apply imperceptibly small perturbations to input data that are nearly undetectable to the human eye, causing a trained deep learning model to produce completely incorrect outputs.
70%UnverifiedCounterexamples generated by standard adversarial attacks are independent and scattered, making it difficult to systematically cover the failure boundary; whereas the counterexample sets generated by the enumeration-guided method are more compact and representative, capturing the failure boundary more precisely.
50%UnverifiedThe adversarial attack phenomenon was first discovered by Szegedy et al. in 2013, after which Goodfellow et al. proposed classic attack methods such as FGSM.
50%