This article mainly aims at B Stand on Teacher Zhang Xiaojian , A summary of the lecture on differential privacy theory and Application , Promoting learning by writing ,B Station video link and PPT It will be put after the text .

1. Differential privacy

1.1 Data release privacy disclosure

Today, , We are in an era of big data . By mastering some of our data , Can optimize the program , Design better products , Improve our quality of life , meanwhile , We also want our data to be used , Can keep privacy from being revealed （ That is, some or some attributes of a particular individual ）, This requires a method of privacy protection .

If the data is not treated with privacy protection , Or if the method is too simple , There is a high probability of privacy disclosure , And then damage your personal life ：

• Carnegie Mellon university LatanyaSweeney Will be anonymous GIC database ( Include the date of birth of each patient 、 Gender and zip code ) When linked to the voter registration record , You can find the medical records of the governor of Massachusetts .
• 2018 Cambridge analytics “ manipulation ” US election Events , Through illegal acquisition 8700 ten thousand Facebook User data , Build psychoanalytic models , Analyze the personality characteristics of Internet users , And then targeted the implementation of false advertising .

1.2 Privacy protection methods

There are current methods to protect privacy , Mainly add the first and third ：

• K- anonymous (K-Anonymous)
• K- Anonymous variant （L-diversity,T-Closeness）
• Differential privacy （differential Privacy）

1.2.1 K- anonymous

The figure above shows the data to be published , It contains three types of data ： Identification data , Quasi identification data and sensitive data

• Identification data ： full name , Data that can uniquely identify the user's identity
• Quasi identification data ： Gender , Zip code , date of birth , Quasi identification data refers to some attributes between identification data and sensitive attributes , These attributes are combined with other data tables （ Link Attack ）, It can also identify the specific information of the patient .
• sensitive data ： disease , The core information of publishing data .

K- Anonymity mechanism ： Data before release , First erase the identification data , Secondly, align the identification data for processing , Make each data at least K-1 Data cannot be distinguished .

The following is an example of 2- Anonymous data , First delete the name , Then on “ Gender , Zip code and date of birth ” Wait for information to be anonymous . As can be seen from the table , Any piece of data cannot be distinguished from at least one other piece of data .

Link Attack ： It refers to the operation of the attacker to obtain and publish data through other channels , To infer private data , Thus causing privacy disclosure , It is equivalent to an expansion of the dimension of personal information .

CDP And LDP The practical application of