[dsn'19] Deep Validation: Toward Detecting Real-world Corner Cases for Deep Neural Networks
Keywords: DL Robustness, AE detection
Takeaways:
Background
1. AE detection
Design
1. Motivation
Legitimate input range/probability distribution for every layer is ill-defined, this Is because:
- the decision functions of these layers are learned on their own rather than manually designed by the developers
- the classification rules they derive from the training data are encoded in millions of parameters, which are nearly impossible to translate
Key observation: images of different classes can fire different patterns and follow different paths when transferred from one area into another one when going through layers
(相同的label應(yīng)該有相近的激活路徑/隱層表示, 不同的label的也不同)
2. Overview
overview
每類每層train一個(gè)OCSVM浮毯，然后用signed distance最后算累計(jì)（求和）誤差，大于一定閾值則判定為corner case
Experimental Results
Personal Response
+ Strengths：
- Weaknesses:
【轉(zhuǎn)載聲明】 轉(zhuǎn)載或引用本博客文章請注明出處 -- AISecPaperShare
