Work

Scalable Vertical Federated Learning for Financial Systems with Prototype Consistency Based Malware and Backdoor Defense Strategy

Year: 2026
Type: article
Abstract: Vertical federated learning (VFL) enables multiple financial institutions to collaboratively train machine learning models without sharing raw data, thereby addressing privacy regulations and competitive barriers. However, the distributed nature of VFL introduces new attack surfaces, particularly backdoor poisoning where malicious participants embed hidden triggers that cause targeted misclassification during inference. Existing defenses often impose significant communication overhead or degrade model accuracy, hindering scalability in large-scale financial ecosystems. This paper proposes a scalable VFL framework augmented by a prototype consistency based backdoor defense strategy that leverages feature-space regularization to detect and mitigate poisoned updates without requiring access to raw client data. The architecture employs a split neural network with a central server maintaining a prototype bank derived from benign training examples; each client’s intermediate representation is compared against these prototypes during aggregation, and updates that deviate beyond adaptive thresholds are filtered or corrected. We discuss the structural trade-offs between defense strength and model utility, the deployment challenges in heterogeneous financial infrastructures, and the governance implications for regulatory compliance. Through a system-level analysis spanning scalability, robustness, fairness, and policy considerations, we demonstrate that prototype consistency offers a lightweight yet effective defense compatible with real-world constraints such as bandwidth limits, asynchronous communication, and non-IID data distributions. The proposed strategy is evaluated against state-of-the-art backdoor attacks and compared with alternative defenses, showing favorable performance in preserving model integrity while maintaining training efficiency. This work contributes to the intersection of secure federated learning, adversarial machine learning, and financial system design, providing a path toward trustworthy AI in regulated industries.
Language: English
Open Access status: diamond
OpenAlex ID: W7220249006
FWCI:
Cites: 18
Cited by:
Related to: 10