抗機器學習攻擊之高穩定物理不可被複製函數設計

Abstract

隨著物聯網的應用越來越廣泛，而物聯網中的安全性的需求也隨之提高，而物理不可被複製函數(physically unclonable function, PUF)，非常適合拿來當物聯網的安全系統，因為每個晶粒(die)都具有其特有的隨機函數，可以使每一個裝置上都裝有物理不可被複製函數的晶片，並先於註冊時期先將認證的挑戰碼和其所對應之回應碼存於雲端伺服器，在認證時由雲端伺服器傳出多筆挑戰碼給所要認證之裝置，並等待回傳對應值，而裝置端需填入所有挑戰碼所對應之回應碼，而錯誤量要少於雲端伺服器所能接受之數量才算認證成功。

由於PUF分為兩類，一類為弱PUF，主要用於產生一組隨機亂數，用於加解密系統所要的亂數值，另外一類為強PUF，主要用於裝置端的身分驗證，論文會直接切入強PUF的部分，並介紹各種現有的強PUF，以及各個電路的優缺點，和目前機器學習攻擊強PUF的演算法，接著介紹用以區別強PUF好壞的重要參數。

最後著重在我們所提出的強PUF，除了在不同的電壓、溫度變化的情況，都足夠穩定度，也具備防機器學習攻擊的能力，甚至是目前面積最小的強PUF。我們使用TSMC 28nm製程來分析其電路行為，然後使用TSMC 28nm製程下線成晶片，未來會做成測試晶片並量測實際結果。

With the rise of the Internet of things, the demand for security in the Internet of things is also increasing. Physically unclonable function (PUF) is suitable to be used as a security system for the Internet of things because each die has its own specific random function. Save the challenge response pairs in the cloud server during the registration period. During authentication, the cloud server sends out several challenges to the device to be authenticated and waits for the corresponding value to be returned. The device needs to fill in the corresponding response to all challenges, and the number of errors should be less than the number accepted by the cloud server before the authentication is successful.

PUF is divided into two categories, one is weak PUF, which is mainly used to generate a set of random numbers; the other category is strong PUF, which is mainly used for device authentication. The paper will introduce various existing strong pufs, as well as the advantages and disadvantages of each circuit. Introduce the current machine learning attack PUF algorithm.

Finally, we focus on the proposed strong PUF, which is stable enough in different voltage and temperature, and has the ability to prevent machine learning attacks. The proposed strong PUF has the smallest area at present. We use the TSMC 28nm technology to analyze its circuit behavior.