[paper]Universal adversarial perturbations
本文提出了計算自然圖像中普遍擾動的算法，在神經網絡上具有很好的泛化性，並且揭示了模型在高維空間中決策邊界的幾何聯繫。並且說明了在輸入空間中存在單個方向的潛在安全漏洞，攻擊者可能會利用這些漏洞造成模型對大多數自然圖像分類錯誤。 universal 算法： μ \mu μ是圖像分佈 v v v是universal擾動 k ^ \hat{k} k^是模型 universal擾動 v v v需要滿足兩個條
相關文章
- 1. Adversarial Perturbations
- 2. Multi-Label Adversarial Perturbations
- 3. 《Universal adversarial perturbations》文獻閱讀筆記
- 4. 對抗機器學習——Universal adversarial perturbations
- 5. 關於Distillation as a Defense to Adversarial Perturbations against Deep Neural Networks的理解
- 6. [paper]Distillation as a Defense to Adversarial Perturbations against Deep Neural Networks
- 7. Data-Free Adversarial Perturbations for Practical Black-Box Attack閱讀筆記
- 8. Paper Review: Adversarial Examples
- 9. 【遷移攻擊筆記】圖像＆模型同時遷移2017Universal adversarial perturbations
- 10. Hidden Voice Attack | Adversarial Attack on Audio
- 更多相關文章...
相關標籤/搜索
每日一句
-
每一个你不满意的现在，都有一个你没有努力的曾经。
歡迎關注本站公眾號,獲取更多信息
相關文章
- 1. Adversarial Perturbations
- 2. Multi-Label Adversarial Perturbations
- 3. 《Universal adversarial perturbations》文獻閱讀筆記
- 4. 對抗機器學習——Universal adversarial perturbations
- 5. 關於Distillation as a Defense to Adversarial Perturbations against Deep Neural Networks的理解
- 6. [paper]Distillation as a Defense to Adversarial Perturbations against Deep Neural Networks
- 7. Data-Free Adversarial Perturbations for Practical Black-Box Attack閱讀筆記
- 8. Paper Review: Adversarial Examples
- 9. 【遷移攻擊筆記】圖像＆模型同時遷移2017Universal adversarial perturbations
- 10. Hidden Voice Attack | Adversarial Attack on Audio