Asian Research Journal of Current Science

Deep learning is becoming increasingly collaborative and outsourced, with neural computation split between resource-constrained clients and capable servers. While split learning reduces the computational load on the client, the transferred intermediate activations are vulnerable to feature-space reconstruction and model inversion attacks, while gradient exchanges are vulnerable to membership inference attacks. In this paper, we design, formalise, and empirically evaluate a tri-layered hybrid privacy architecture that integrates split learning, differential privacy (DP), and Cheon-Kim-Kim-Song (CKKS) fully homomorphic encryption (FHE) into a single coherent pipeline. We demonstrate how each layer addresses a different threat class: split learning restricts exposure of raw data, DP bounds statistical leakage from gradients and activations, and CKKS enforces computational indistinguishability for data that must cross the network boundary. We provide a formal composition of the privacy layers, a reference architecture with client- and server-side responsibilities, and an experimental evaluation across three datasets, including a seven-class electrical fault-diagnosis dataset. Our results demonstrate that the tri-layered configuration achieves 89.9% accuracy at ϵ = 8 (only 4.7 percentage points below a non-private baseline of 94.6%), with computational security against an honest-but-curious server and a measured latency overhead of ≈10.4 × relative to plaintext split inference. We also quantify the marginal contribution of each layer to overall threat-model coverage and provide guidance on jointly tuning the noise multiplier, clipping norm, and CKKS multiplicative depth rather than treating them independently.