HoloCode: Hybrid Optical-Electronic Edge Encoding for Privacy-Preserving Cloud Training
Abstract
Privacy-preserving machine learning defends against adversaries without sacrificing task accuracy. In latency-critical, resource-constrained settings, existing cryptographic and encoding approaches incur heavy overheads, causing intolerable delays and energy costs. We present HoloCode, a hybrid optical–electronic pipeline delivering strong privacy with sub-5ms latency at a fraction of state-of-the-art energy. HoloCode encodes task-relevant signals, shields sensitive features, resists inversion attacks, and locks models with a private key preventing misuse. It builds on an edge–cloud framework pushing inference to the edge to cut latency, at the cost of higher edge energy. To break this, we leverage zero-energy optical processing to reduce latency and energy simultaneously. Against strong baselines, HoloCode achieves 10× faster inference and 50% lower edge energy, preserving accuracy while resisting leakage and reconstruction attacks.