基於CNN方法之真假人臉識別模型;SPOOF FACE DETECTION USING CONVOLUTIONAL NEURAL NETWORKS

NCU Institutional Repository > 資訊電機學院 > 資訊工程研究所 > 博碩士論文 > Item 987654321/86734

jsp.display-item.identifier=請使用永久網址來引用或連結此文件: https://ir.lib.ncu.edu.tw/handle/987654321/86734

题名:	基於CNN方法之真假人臉識別模型;SPOOF FACE DETECTION USING CONVOLUTIONAL NEURAL NETWORKS
作者:	賴冠妤;Lai, Kuan-Yu
贡献者:	資訊工程學系
关键词:	真假人臉辨識;SPOOF FACE DETECTION
日期:	2021-08-23
上传时间:	2021-12-07 13:10:20 (UTC+8)
出版者:	國立中央大學
摘要:	在2020年，席捲全球的COVID-19病毒，使得許多國家面臨封城的狀況，因此人們得在家工作並依靠3C產品來完成日常所需的採買，然而在大量依靠網路的生活下，重要資訊傳播的安全性正面臨挑戰，生物辨識成為最新的安全保護趨勢。生物辨識乃根據生物的生理或行為特徵(如:臉部、虹膜、聲紋、指紋、簽名等)來辨識身分，其中臉部辨識具有唯一性，且手機就可以拍攝臉部，所以成本低廉，再者臉部在取像時，並不需要直接接觸，這在疫情肆虐的險峻時刻是很重要的一項優勢，然而有很多方法會使臉部辨識系統失靈，例如相片或影片攻擊。常見的臉部辨識攻擊— Presentation Attacks（PA） [1]，其主要目的是冒充他人的身分或者避免被系統辨識，它主要包含print attack(如相片、列印出的圖片)、replay attack(播放影片)、3D mask attack(如戴面具、3D列印面具)；判斷臉部仿冒攻擊的難易度，由易到難依序為：相片攻擊、影片重播攻擊、列印攻擊、3D mask攻擊，其中以print attack和replay attack是最常見的攻擊。我們通過使用相片以及列印出的圖片來做為假臉數據集，對CNN的不同結構進行比較研究，最終發現儘管CNN模型為最簡單的神經網路，但他擊敗了大多數複雜模型。我們的研究結果顯示，以5~8層的CNN模型可以達到99%的準確率及召回率，並以6、7層CNN模型得到最佳的Performance。 ;In 2020, Coivd-19 has swept the world and let lots of countries be on lockdown. Therefore, people have to stay at home and rely on 3C products to complete their daily purchases. However, since relying heavily on the Internet, the security of the dissemination of important information is facing challenges. Biometric identification has become the latest security protection trend. Biometric recognition is based on biological or behavioral characteristics (such as: face, iris, voiceprint, fingerprint, signature, etc.) to recognize identity. Face recognition is unique, and mobile phones can capture faces, so the cost can be low. Furthermore, the face does not need to be in direct contact when capturing images. This is a great advantage in this precarious time. However, there are still many ways to make the facial recognition system fail, such as print attack or replay attack. A common face recognition attack — Presentation Attacks (PA) [1], whose main purpose is to impersonate others or avoid being recognized by the system. It mainly includes print attack (such as photos, printed pictures), replay attack (playing video), 3D mask attack; judging the difficulty of a spoof face attack, from easy to difficult: photo attack, video replay attack, print attack, 3D mask attack. Thus, print attack and replay attack are the most common attacks. We compared the different structures of CNN by using photos and printed pictures as the spoof face dataset, and finally found that although the CNN model is the simplest neural network, it beats most complex models. Our research shows that a CNN model with 5 to 8 layers can achieve 99% accuracy and recall, thus 6 and 7 layers CNN model can get the best performance.
显示于类别:	[資訊工程研究所] 博碩士論文

文件中的档案:

档案	描述	大小	格式	浏览次数
index.html		0Kb	HTML	106	检视/开启

在NCUIR中所有的数据项都受到原著作权保护.

社群 sharing

数据加载中.....