Centrala begrepp
Adversarial example soups improve transferability by averaging multiple batches of adversarial examples under different hyperparameter configurations.
Sammanfattning
Adversarial examples crafted on surrogate models mislead target models effectively.
Existing methods focus on input, gradient, and features, neglecting the adversarial examples themselves.
Adversarial example soups enhance transferability without additional generation time.
Three types of soups: mixup, uniform, and combination, improve attack success rates.
Experiments show AES outperforms baseline attacks on ImageNet dataset.
Statistik
"Extensive experiments on the ImageNet dataset show that our methods achieve a higher attack success rate than the state-of-the-art attacks."
Citat
"Averaging multiple batches of adversarial examples under different hyperparameter configurations, referred to as 'adversarial example soups', can often enhance adversarial transferability."