Adaptive Neural Architecture Search for Cryptographically Efficient Private Inference Under Latency and Accuracy Constraints

Abstract

Private inference has become a central concern for machine learning systems that process sensitive user data across organizational, jurisdictional, and infrastructural boundaries. Cryptographic techniques such as secure multi-party computation and homomorphic encryption can preserve confidentiality during inference, but they impose substantial computational and communication overheads that interact poorly with conventional deep neural network design. This paper examines adaptive neural architecture search as a system-level mechanism for producing cryptographically efficient private inference pipelines under explicit latency and accuracy constraints. Rather than treating architecture search as a model-centric optimization problem, the analysis frames it as a socio-technical infrastructure capability that links model topology, cryptographic protocol selection, deployment topology, governance, and lifecycle sustainability. The paper argues that adaptive search methods can reduce non-linear activation cost, communication rounds, memory pressure, and operational uncertainty, but only when they are constrained by deployment realities and accountable design practices. It further considers robustness, fairness, policy implications, and cross-domain comparisons, including healthcare, finance, and public-sector applications. The conclusion is that cryptographically efficient private inference requires not merely better search algorithms but also institutional mechanisms that make latency, accuracy, privacy, and sustainability trade-offs visible, auditable, and governable over time.