The Synthetic Threat: AI-Generated Identity Fraud, Strategic Risk Management, and Deep Learning in U.S. Banking.

Abstract

Generative artificial intelligence has reduced the cost of producing realistic identity documents, facial imagery, voice samples, and supporting application narratives. These capabilities intensify a long-standing U.S. banking problem: synthetic identities that combine genuine and fabricated attributes and mature over time before a credit bust out or account takeover. This paper develops a privacy-preserving deep learning architecture for detecting AI-generated identity fraud during remote onboarding and early account activity. The design transfers three source-derived elements: Hybrid Masked Encryption (HME) for protected data aggregation, an Intel SGX-style Trusted Execution Environment (TEE) for attested enclave processing, and a five-layer cybersecurity model connecting data resources, preprocessing, intelligent analytics, decision orchestration, and governance. A multimodal detector is then proposed to fuse document-forensics, selfie-video, liveness, application, credit-file, device, network, and behavioral signals. The source HME study reports 25 ms encryption and 20 ms decryptions for a 1 MB dataset, while the TEE study offers proof-of-concept architecture using remote attestation, local attestation, and sealed storage. These findings support technical feasibility but do not establish fraud-detection accuracy in banking. Accordingly, this paper defines a temporal and generator-holdout validation protocol centered on precision-recall performance, recalling operational false-positive constraints, calibration, demographic error analysis, latency, and privacy leakage. The principal contribution is an auditable blueprint that joins fraud intelligence with data protection, while preserving a clear boundary between source evidence and results that still require empirical testing.