Research into Developing a Framework for Cross-Organizational Financial Information Confidentiality Safeguarding and Cooperative Examination via a Distributed Learning Structure

Abstract

Owing to the inherent contradiction between "data silos" and the threat of financial information leakage within cross-firm cooperative audits, this research introduces a framework for fiscal data confidentiality preservation and joint auditing via federated learning. This architecture incorporates a blockchain consensus protocol that facilitates dual-factor participation, utilizes differential privacy to shield the integrity of intermediate weights during federated training, and develops a dynamic model synthesis strategy rooted in quality metrics and trust ratings; the operation of this procedure is autonomously initiated by smart contracts. A collective-based synthesis auditing mechanism is employed to swiftly pinpoint corporate participants who have submitted tainted data, while the collective financial analytics framework and equity among data contributors remain accurate. Based on the findings, the precision of the universal model under defense remains nearly identical to that of the non-poisoned group-summarized model, reaching 91.39%. The effectiveness of the universal model in poisoning scenarios is 90.82%. Free-rider maneuvers are more conspicuous and simpler to identify in collective summation, thereby preventing the exposure of transient parameters, and consequently, a reliable federated learning structure with robust privacy defense is realized.