ByteZip: Efficient Lossless Compression for Structured Byte Streams Using DNNs

Data compression plays a key role in efficient data storage, transmission, and processing. With the fast development of deep learning techniques, deep neural networks have been used in this field to achieve a higher compression rate. Deep learning-based general-purpose lossless compression techniques are formulated as an autoregressive sequential prediction problem. These methods are state-of-the-art in terms of compression ratio but not practical due to runtime and resource constraints. Recent advances in lossless image compression using non-autoregressive methods for probability modeling prove to be a faster and more practical approach. In this paper, we propose ByteZip, a lossless compression method based on the non-autoregressive approach for known or defined structured byte streams. ByteZip involves hierarchical probabilistic modeling using autoencoders and density mixture models. This approach reduces the overhead of sequential processing. The goal is to design a practical lossless compressor with faster compression and decompression along with a competitive compression ratio. Experiments show that the proposed approach achieves a 64× higher compression speed than the state-of-the-art transformer-based model TRACE with an overhead of only 5% less size reduction on average. Our approach outperforms general-purpose compressors such as Gzip (23% more size reduction on average) and 7z (16% more size reduction on average).