A new secure approach for AI-based compression across various domains

Abstract

Data security is a fundamental objective of all information systems, especially those handling financial data, which involves technologies, strategies, and measures to defend sensitive financial data. Consequently, data compression is regarded as an effective technique for reducing storage requirements for all generated and stored data. Therefore, it has become necessary to design an integrated system that provides security for financial data in addition to preserving the required storage space as much as possible. Accordingly, the proposed approach consists of three main components: the sender, the receiver, and the graphical user interface (GUI). The system accepts two types of input documents (i.e., images and PDF files), extracts their data into a structured format, and then classifies the processed data into sensitive and non-sensitive categories. Watermarking and encryption are subsequently applied only to the sensitive data to ensure privacy preservation and enhance overall security. The most suitable compression algorithm was selected based on the results obtained when comparing standard algorithms (Zstd, LZMA, and Brotli). The results showed that the proposed system, Autoencoder and ZSTD, which achieved compression efficiency improvements ranging from 26.4% to 52.8% compared to the traditional ZSTD algorithm across three different-sized datasets (6,784 bytes, 169,668 bytes, and about 327,000 bytes), while maintaining high entropy values ​​ranging from 5.83 to 5.99 bits/byte. Furthermore, chi-square tests showed high values, reaching 11,015, confirming the high level of statistical randomness and security of the resulting data.