Sharing Cyber Threat Warnings Across Banks in Real Time Without Exposing Private Data: A Privacy-Preserving Federated Threat Intelligence Framework
Main Article Content
Abstract
Banking databases contain sensitive customer, transaction, payment, credit, identity, and regulatory information. Conventional database access-control mechanisms can determine whether a user has permission to access a resource, but they are less effective when an authorized employee, administrator, contractor, or compromised account uses legitimate privileges for an unauthorized purpose. Machine-learning systems can identify unusual behavior, although their predictions are probabilistic and may be difficult to explain or enforce consistently. This paper proposes a neuro-symbolic artificial intelligence framework for policy-aware database access control and insider-threat detection in banking systems. The framework combines neural behavioral models with symbolic authorization rules, attribute-based policies, separation-of-duty requirements, business-purpose restrictions, data classifications, and transaction context. The neural component analyzes database queries, authentication activity, access frequency, record volume, temporal patterns, device posture, and peer-group behavior. Its outputs are transformed into risk predicates that are evaluated by a symbolic policy engine. The engine produces graduated decisions, including permit, permit with masking, result limitation, step-up authentication, human review, temporary suspension, and denial. The proposed architecture preserves deterministic security rules while allowing access decisions to respond to changing behavioral risk. It also creates an explanation trace that records the policies, attributes, risk indicators, and evidence associated with each decision. A design-science methodology, formal decision model, banking use cases, threat analysis, and evaluation framework are presented. The paper argues that neuro-symbolic integration can improve continuous authorization, contextual insider-threat detection, explainability, and regulatory auditability. However, the system should support human security decision-making rather than independently determine employee guilt or impose disciplinary action.