µ¥ÀÌÅͺ£À̽º ¿¬±¸È¸Áö(SIGDB)
Current Result Document :
ÇѱÛÁ¦¸ñ(Korean Title) |
½ÉÃþ½Å°æ¸Á¿¡¼ ÇнÀ ¿Ü ºÐÆ÷ µ¥ÀÌÅÍ Å½Áö ¾Ë°í¸®Áò |
¿µ¹®Á¦¸ñ(English Title) |
Detecting Out-of-distribution data for Deep Neural Networks |
ÀúÀÚ(Author) |
±è»ó¿ø
ÀÌÈ£ÁØ
±è¼ºÈÆ
¼Áö¿ø
Sangwon Kim
Hojun Lee
Seonghun Kim
Jiwon Seo
|
¿ø¹®¼ö·Ïó(Citation) |
VOL 35 NO. 02 PP. 0066 ~ 0077 (2019. 08) |
Çѱ۳»¿ë (Korean Abstract) |
½ÉÃþ½Å°æ¸Á ȤÀº ¿¡¼ÀÇ ÇнÀ ¿Ü ºÐÆ÷ µ¥ÀÌÅÍ (Deep Neural Network, DNN) (Out-of-Distribution Data)¶õ ÇØ´ç ½ÉÃþ½Å°æ¸ÁÀ» ÇнÀ½ÃŲ µ¥ÀÌÅÍ¿Í »ó´çÇÑ Â÷À̸¦ º¸ÀÌ´Â µ¥ÀÌÅ͸¦ ¸»ÇÑ´Ù. ÇÏÁö¸¸, ½ÉÃþ½Å°æ¸ÁÀÌ µµÃâÇÏ´Â °á°ú¿¡ ´ëÇÑ ±Ù°Å¿Í µµÃâ ¹æ¹ýÀ» ÀÌÇØÇÏ´Â °ÍÀÌ ¾î·Æ±â ¶§¹®¿¡ ÇнÀ ¿Ü ºÐÆ÷ µ¥ÀÌÅ͸¦ ŽÁöÇϱ⠾î·Æ´Ù. º» ³í¹®¿¡¼´Â ¸¹Àº µ¥ÀÌÅÍµé ¼Ó¿¡¼ ±×·¯ÇÑ ÇнÀ ¿Ü ºÐÆ÷ µ¥ÀÌÅ͸¦ Á¤È®ÇÏ°Ô ½Äº°Çϱâ À§ÇØ µ¶Ã¢ÀûÀÎ ±â¹ýÀ» Á¦¾ÈÇÑ´Ù ÀÌ ±â¹ýÀº ÇнÀ ³» ºÐÆ÷ µ¥ÀÌÅÍ¿¡ ´ëÀÀÇÏ´Â Àº´Ð ´º·ÐµéÀÌ Æ¯Á¤ÇÑ ºÐÆ÷¸¦ °¡Áø´Ù´Â °¡Á¤À» ±â¹ÝÀ¸·Î ÇÑ´Ù. ÀÌ·¯ÇÑ, °¡Á¤ ÇÏ¿¡¼ º» ±â¹ýÀº ÇнÀ ³» ºÐÆ÷ µ¥ÀÌÅ͸¦ ºñ½ÁÇÑ Àº´Ð ´º·Ð ÆÐÅÏÀ» °¡Áø °Í³¢¸® ±ºÁýÈ ÇÑ ÈÄ ±× °á°ú¸¦ ¹ÙÅÁÀ¸·Î ÇнÀ ¿Ü ºÐÆ÷ µ¥ÀÌÅ͸¦ °ËÃâÇÑ´Ù |
¿µ¹®³»¿ë (English Abstract) |
The inference outcome of a Deep Neural Network (DNN) may be incorrect if the input is substantially different from the distribution of its training data. Thus detecting such out-of-distribution (OOD) data is essential for a DNN to be securely applied. Because we do not yet understand why and how a DNN makes an inference, it is hard to identify OOD data by examining the inference process. We propose a novel method to accurately detect OOD data and in-distribution (ID) data. Our method assumes that there exists a specific distribution of the values of hidden neurons for ID data. Under this assumption, our method clusters ID data with similar hidden-neuron patterns and detect OOD data based on the clustering results. |
Å°¿öµå(Keyword) |
Deep Neural Network
Out-of-Distribution detection
Deep Neural Network Security
½ÉÃþ½Å°æ¸Á ÇнÀ
ºÐÆ÷ µ¥ÀÌÅÍ Å½Áö
½ÉÃþ½Å°æ¸Á º¸¾È
|
ÆÄÀÏ÷ºÎ |
PDF ´Ù¿î·Îµå
|