ÇѱÛÁ¦¸ñ(Korean Title) |
ActiveBoostThief: ½Å·ÚÇÒ ¼ö ÀÖ´Â ´Éµ¿Àû ÇнÀÀ» ÀÌ¿ëÇÑ ¸ðµ¨ ÃßÃâ °ø°Ý |
¿µ¹®Á¦¸ñ(English Title) |
ActiveBoostThief: Model Extraction Attack Using Reliable Active Learning |
ÀúÀÚ(Author) |
³²¿µÀº
°ÁØÇõ
ÀÌÀç±æ
Youngeun Nam
Junhyeok Kang
Jae-Gil Lee
|
¿ø¹®¼ö·Ïó(Citation) |
VOL 48 NO. 01 PP. 0594 ~ 0596 (2021. 06) |
Çѱ۳»¿ë (Korean Abstract) |
|
¿µ¹®³»¿ë (English Abstract) |
As machine learning models are being applied in practice, the security for models is becoming more significant. A model extraction attack, one of the types of adversarial attack, exploits the open application programming interfaces (APIs) to figure out the object model. Prior works for model extraction using active learning have a lack of confidence in the thief model when selecting the instances for queries. We propose ActiveBoostThief framework that accomplishes a model extraction using active learning to complement the model¡¯s reliability. We demonstrate with experiments that our model improves the reliability of the substitute model and enables more accurate performance in the model extraction attack compared with the existing state-of-the-art baseline, ActiveThief. |
Å°¿öµå(Keyword) |
|
ÆÄÀÏ÷ºÎ |
PDF ´Ù¿î·Îµå
|