Privacy-Preserving Artificial Intelligence in Enterprise Applications: The Role of Federated Learning and Differential Privacy in Organizational Data Protection

Abstract

Enterprise AI systems increasingly run on personal, financial, and clinical records, which leaves organizations balancing analytical value against privacy exposure and regulatory liability. This paper reviews the two privacy-enhancing technologies (PETs) that dominate that balance in practice, federated learning (FL) and differential privacy (DP), and asks how they behave once they leave the laboratory and enter a governed production environment. Using an adapted systematic-review protocol, we synthesize 28 primary and secondary sources published between 2006 and 2024 across five dimensions: conceptual foundations, architectural design, adversarial threats, complementary cryptographic techniques, and enterprise deployments. FL removes the need for a central repository of raw records but supplies no formal guarantee on its own; DP supplies an (ε, δ) guarantee but spends model accuracy to buy it. Neither is sufficient alone. We argue that the decision that matters most in an enterprise setting is rarely FL versus DP but rather the unit of privacy and the placement of trust, and that this pair of choices, more than the headline value of ε, determines both the utility cost and the compliance argument an organization can defend. The paper develops a four-layer taxonomy positioning FL and DP alongside secure aggregation, homomorphic encryption, and trusted execution, and maps each layer onto specific obligations under the GDPR, the CCPA as amended by the CPRA, and HIPAA. Residual risks such as membership inference, gradient inversion, and backdoor insertion survive both techniques, and managing them is a governance problem as much as an engineering one.