基於全同態加密車載系統之兒童遺留偵測

Abstract

全同態加密(FullyHomomorphicEncryption) 是能夠進行密文運算的加密方法。全同態加密能夠保護線上服務的使用者隱私，讓使用者能夠放心地將資訊上傳至雲端進行各式應用像是匿名投票、去中心化身份與機器學習。其中機器學習的應用在由於硬體上的限制，因此在推論時需要大量的時間，如何降低推論所需要的時間，在現在是一大挑戰。 先前研究使用了全同態加密中TFHE(FastFully Homomorphic Encryption Scheme Over the Torus) 的演算法實現VGG9機器學習模型，並且對於CIFAR10進行分類，本研究主要是針對車載系統上的影像使用全同態加密進行分類，使用不同的機器學習模型來判斷並分析準確度以及推論所需的時間，改善VGG9模型，在維持一定的準確度情況下，大幅節省推論所需的時間，由於現階段對於全同態加密還沒有特殊應用積體電路，因此僅能夠實現較小的模型，但是在未來有特殊應用積體電路出現後，可以參照本篇論文提出的各種優化方法來優化更大的神經網路模型。

Fully Homomorphic Encryption (FHE) is an encryption method that allows for operations on encrypted data. It protects the privacy of users of online services, enabling the secure upload of information to the cloud for various applications, such as anonymous voting, decentralized identity, and machine learning. One of the challenges, particularly in machine learning applications, is the significant amount of time required for inference due to hardware limitations. Previous research implemented a VGG9 machine learning model using the TFHE (Fast Fully Homomorphic Encryption Scheme Over the Torus) algorithm to classify CIFAR10. This study focuses on using fully homomorphic encryption to classify images in smart vehicles. Different machine learning models are used to evaluate and analyze accuracy and inference time, aiming to improve upon the VGG9 model by significantly reducing inference time while maintaining a certain level of accuracy. Due to the current lack of specialized application-specific integrated circuits (ASICs) for fully holomorphic encryption, only smaller models can be realized. However, when such ASICs become available in the future, the various optimization methods proposed in this paper can be applied to larger neural network models.