Extraction of trigger and mask from poisoned data using modified Activation Clustering and Neural Cleanse methods

Lozinskii, I.; Kostyumov, V.; Stroeva, E.

Авторы: Lozinskii Ivan, Kostyumov Vasily, Stroeva Ekaterina
Журнал: International Journal of Open Information Technologies
Том: 11
Номер: 7
Год издания: 2023
Издательство: Лаборатория Открытых Информационных Технологий факультета ВМК МГУ им. М.В. Ломоносова
Местоположение издательства: Москва
Первая страница: 1
Последняя страница: 10
Номер статьи: 1
Аннотация: In some works, it has been repeatedly noticed that the popular Neural Cleanse method does a poor job of restoring triggers and masks that occupy a significant part of the image, since the method looks for the least poisoning change. To solve this problem, we proposes a method for extracting a trigger from an averaged image of a poisoned cluster of images. The trigger can be extracted by filtering the pixel color intensity of the averaged image. To select clusters of images, a modification of the Activation Clustering method is used. The experiments were conducted on data from the Trojan Detection Challenge, NeurIPS 2022. In this data, a single trigger translates any image to the target class. In such a poisoning model, the original Activation Clustering shows poor results. So we proposed a modified version of the Activation Clustering in the article. To restore the mask to the selected trigger, a modification of the Neural Cleanse method was developed. The developed method shows a significantly higher quality of trigger isolation in comparison with the original Neural Cleanse.
Добавил в систему: Строева Екатерина Николаевна

	ИСТИНА	Войти в систему Регистрация
	ИСТИНА ЦЭМИ РАН
	Главная Поиск Статистика О проекте Помощь

ИСТИНА

ИСТИНА ЦЭМИ РАН

Extraction of trigger and mask from poisoned data using modified Activation Clustering and Neural Cleanse methodsстатья Исследовательская статья

Прикрепленные файлы