Method and apparatus of revising a deep neural network for adversarial examples

A device for calibrating a deep neural network comprising a plurality of neurons according to one embodiment may comprise: an input part configured so as to input-receive data comprising noise and normal data; a calculation part configured so as to calculate an activation frequency of a neuron activ...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: SON SOOEL, KO GIHYUK, LIM GYUMIN, LEE SUYOUNG
Format: Patent
Sprache:eng ; kor
Schlagworte:
Online-Zugang:Volltext bestellen
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:A device for calibrating a deep neural network comprising a plurality of neurons according to one embodiment may comprise: an input part configured so as to input-receive data comprising noise and normal data; a calculation part configured so as to calculate an activation frequency of a neuron activated by the data comprising the noise and an activation frequency of a neuron activated by the normal data among the plurality of neurons; and a calibration part configured so as to calibrate the deep neural network based on the activation frequency of the neuron activated by the data comprising the noise and the activation frequency of the neuron activated by the normal data. Therefore, the present invention can efficiently defend against adversarial examples. 일 실시예에 따른 복수개의 뉴런들을 포함하는 심층신경망을 보정하는 장치는, 노이즈를 포함하는 데이터와 정상 데이터를 입력받도록 구성되는 입력부, 상기 복수개의 뉴런들 중 상기 노이즈를 포함하는 데이터에 의해 활성화되는 뉴런의 활성화 빈도수와 상기 정상 데이터에 의해 활성화되는 뉴런의 활성화 빈도수를 계산하도록 구성되는 계산부, 및 상기 노이즈를 포함하는 데이터에 의해 활성화되는 뉴런의 활성화 빈도수와 상기 정상 데이터에 의해 활성화되는 뉴런의 활성화 빈도수를 기반으로 상기 심층신경망을 보정하도록 구성되는 보정부를 포함할 수 있다.