Paper 2024/559

Convolution-Friendly Image Compression with FHE

Abstract

During the past few decades, the field of image processing has grown to cradle hundreds of applications, many of which are outsourced to be computed on trusted remote servers. More recently, Fully Homomorphic Encryption (FHE) has grown in parallel as a powerful tool enabling computation on encrypted data, and transitively on untrusted servers. As a result, new FHE-supported applications have emerged, but not all have reached practicality due to hardware, bandwidth or mathematical constraints inherent to FHE. One example is processing encrypted images, where practicality is closely related to bandwidth availability. In this paper, we propose and implement a novel technique for FHE-based image compression and decompression. Our technique is a stepping stone towards practicality of encrypted image-processing and applications such as private inference, object recognition, satellite-image searching or video editing. Inspired by the JPEG standard, and with new FHE-friendly compression/decompression algorithms, our technique allows a client to compress and encrypt images before sending them to a server, greatly reducing the required bandwidth. The server homomorphically decompresses a ciphertext to obtain an encrypted image to which generic pixel-wise processing or convolutional filters can be applied. To reduce the round-trip bandwidth requirement, we also propose a method for server-side post-processing compression. Using our pipeline, we demonstrate that a high-definition grayscale image ($$) can be homomorphically decompressed, processed and re-compressed in $$$$s with a compression ratio of 100/34.4 on a standard personal computer without compromising on fidelity.