当前位置:网站首页>Study notes of deep learning (8)

Study notes of deep learning (8)

2022-04-23 08:40:00 Bai Yanling

This blog post mainly introduces Adversarial Attack.

Motivation

 Insert picture description here

example of attack

Add noise to the matrix of photos (attacked image)
 Insert picture description here
 Insert picture description here
The two photos are different
 Insert picture description here

How to attack

 Insert picture description here
There are many ways to calculate the distance between pictures
among d ( x 0 , x ) d(x^0,x) d(x0,x)need to consider human perception.
 Insert picture description here

attack approach

 Insert picture description here
 Insert picture description here

white box v.s. black box

 Insert picture description here
Black box attack in non-targeted It's easier to succeed
Some people think ,data It's the cause of adversarial attack The culprit of .
Other types of data may also be attack
 Insert picture description here

attack in the physical world

 Insert picture description here
License plate recognition system :
 Insert picture description here

adversarial reprogramming

 Insert picture description here

“backdoor” in model

It's possible to add... To the training materials attack
 Insert picture description here

Defense

passive defense

 Insert picture description here
Slightly blurred , that will do defense,attack The signal of success is special , It could be some kind of... In one direction .
 Insert picture description here
But when the fuzziness is serious , It may also produce some kind of side effect
 Insert picture description here
Another kind defense Approach is to randomization
 Insert picture description here

proactive defense

 Insert picture description here

版权声明
本文为[Bai Yanling]所创,转载请带上原文链接,感谢
https://yzsam.com/2022/04/202204230759110645.html