Homomorphic Encryption for Audio Conferencing

Abstract: Homomorphic encryption (HE) enables computations on encrypted data making it possible to perform calculations on a server without sharing the secret key for decryption. HE is relatively new and has seen significant improvements in recent years in terms of speed of encryption, decryption, operations, and the number of operations possible to perform in a row without damaging the ciphertext. These improvements open up new possibilities, such as using HE in voice over IP (VoIP) and still being able to mix audio streams at the server without decrypting them, creating a scalable and secure application.  This thesis evaluates the performance of three HE schemes, BGV, BFV, and CKKS, implemented by the open-source library Palisade and compares their performance to a non-HE scheme AES in terms of encryption time, decryption time and end-to-end latency. Furthermore, the performance of mixing with HE-encrypted audio files is evaluated. The work is a proof-of-concept that allows the user to send audio files of different sample rates and batch sizes to compare the performance. It also investigates the Quality of Experience by measuring the audio quality after the mixing, using PESQ and POLQA. The results show that BGV performs almost as good as AES for higher batch sizes, while BFV performs almost as good in most cases and trails behind in others. CKKS, on the other hand, is the slowest scheme but is still fast enough in some cases to be considered a possible encryption scheme. Furthermore, the tests show that the HE schemes do not decrease the listening quality much according to PESQ and POLQA, as all tests concluded a result that is considered good or excellent. The number of files being mixed had a negative impact on narrowband and wideband audio streams, while it did not affect superwideband and fullband.