Designing a universal homomorphic instruction set and applications to transciphering with TFHE

Overview & Synopsis

With the developpement of the internet and the increasing flow of communications it facilitates, maintaining the confidentiality of transmitted data remains a significant concern for user privacy. Until now, confidentiality has primarily been ensured during the transmission and, in some cases, the storage of our data on remote servers. However, when we request a distant server to perform calculations on our data (for example, applying a filter to a photo), the server must decrypt the data to perform the desired operations. This presents a vulnerability, as an attacker who gains access to the servers of such service providers could potentially access unprotected user data.One solution to address this security flaw is homomorphic encryption. This innovative type of encryption allows for calculations to be conducted directly on encrypted data, meaning there is no need to decrypt it in order to perform calculations. As a result, our data remains protected throughout the entire process!Despite its promising nature, the deployment of homomorphic encryption still faces several challenges. For instance, homomorphic ciphertexts are known to be very large (often several thousand bytes for just a few bits of information), which impacts both the execution time of calculations and the memory required for their proper functioning.In this thesis, we explore a process called "transciphering," which reduces the bandwidth needed for transmitting homomorphic data, thereby also decreasing latency when sending this data. In our work, we especially focus on block ciphers such as AES.Additionally, we propose a universal instruction set for a homomorphic processor working on 8-bit encrypted inputs. This means we developed an instruction set that directly operates over ciphertexts. Our processor can execute any algorithm with latencies that are competitive with the state of the art.In summary, we propose improvements to the TFHE homomorphic cryptosystem by examining various transciphering methods and block ciphers. We also design a universal instruction set that outperforms other traditional approaches while facilitating the evaluation of any algorithm over encryptions of bytes.