8th IEEE International Workshop on Information Forensics and Security
The increasing advance of Cloud-based solutions brings about serious privacy problems when outsourcing images for their processing in untrusted environments. One of the fundamental privacyaware image manipulations that can be outsourced is denoising, an ubiquitous signal processing primitive with a broad set of applications. Traditional Signal Processing in the Encrypted Domain solutions cannot efficiently address this problem, as they require interactive protocols in order to cope with polynomial operations and comparisons at the same time. We propose methods based on 2-RLWE (Ring Learning with Errors) to efficiently perform the whole image denoising operation on encrypted images in a fully non-interactive way; we show how to combine homomorphic polynomial operations and thresholding without involving decryption or interaction, therefore enabling fully unattended encrypted image processing. We evaluate our solutions for real image sizes and strict security parameters, showing their practicality and feasibility.