Thông tin chung

  English

  Đề tài NC khoa học
  Bài báo, báo cáo khoa học
  Hướng dẫn Sau đại học
  Sách và giáo trình
  Các học phần và môn giảng dạy
  Giải thưởng khoa học, Phát minh, sáng chế
  Khen thưởng
  Thông tin khác

  Tài liệu tham khảo

  Hiệu chỉnh

 
Số người truy cập: 74,204,344

 White-box attack on Google Machine Learning System
Tác giả hoặc Nhóm tác giả: Dang Duy Thang and Toshihiro Matsui
Nơi đăng: The 13th International Workshop on Security, Japan; Số: xxx;Từ->đến trang: xxx;Năm: 2018
Lĩnh vực: Công nghệ thông tin; Loại: Báo cáo; Thể loại: Quốc tế
TÓM TẮT
ABSTRACT
Machine Learning algorithms now achieve state-of-the-art performance on many tasks, including object recognition, face recognition, language translation. Most modern machine learning systems are used in production, can potentially be understood as applied function approximation. That approximation is a mapping from an input x to an output y by learning a training dataset containing several examples of input data and their corresponding outputs. A recent machine learning system is Inception - Google Machine Learning has been trained with millions of labeled images in ImageNet, can classify images as computer mouse, airplanes, container ships or more complex concepts with approximate to human-level performance. However, most machine learning systems are designed based on the assumption that testing data is pulled out from the same distribution of the training data. And this assumption can be violated by adversarial attacks. This phenomenon can be exploited by adversaries capable of crafting inputs. In this work, we show that we can completely fool a Google machine learning system by carefully-crafted perturbations on input data. We create an adversarial example x* by perturbing an original input x. The approach for crafting x* is to solve for the x* that causes the most expected loss, subject to a constraint on the maximum allowable deviation from the original input x. Our result illustrates the Google machine learning system misclassifies images with 100% confidence.
© Đại học Đà Nẵng
 
 
Địa chỉ: 41 Lê Duẩn Thành phố Đà Nẵng
Điện thoại: (84) 0236 3822 041 ; Email: dhdn@ac.udn.vn