Compiler-Assisted Optimization of Privacy-Preserving Neural Network Architectures

Abstract

Privacy-preserving neural network inference has moved from a narrow cryptographic curiosity to a systems problem that spans compilers, runtime schedulers, cryptographic protocols, hardware accelerators, and governance regimes. The central challenge is not simply whether a neural network can be evaluated without exposing inputs or weights, but whether such evaluation can be performed with acceptable latency, cost, accuracy, robustness, and accountability. Compiler-assisted optimization offers a promising response by treating privacy constraints as first-class properties of the computation graph. Instead of applying generic neural network optimizations and then fitting them into a cryptographic protocol, compiler-assisted methods can restructure architectures, lower operator representations, reduce communication rounds, and co-schedule secure computation primitives according to deployment context. This paper examines that emerging design space from a system-level perspective. It discusses how optimizing compilers can bridge secure multi-party computation, homomorphic encryption, and trusted execution environments with neural architecture search, quantization, activation reduction, operator fusion, and memory planning. The analysis emphasizes structural trade-offs rather than narrow benchmark gains. Particular attention is given to the ways privacy-preserving compilation affects fairness, robustness, sustainability, auditability, and policy compliance. The paper argues that compiler-assisted optimization should be understood as an infrastructural and governance instrument, not merely a performance technique. Its long-term value depends on whether it can produce transparent, verifiable, and deployable systems that preserve privacy without shifting unacceptable costs onto operators, users, or the public.