Latent-Space Image Encryption for Edge AI: A Lightweight Autoencoder–Cipher Framework for Secure IoT Vision Analytics
Main article
Abstract
Vision-enabled Internet-of-Things (IoT) deployments increasingly require image confidentiality at the sensor edge, yet conventional block ciphers such as AES impose latency and energy costs that are difficult to absorb on constrained devices. This paper proposes LSE-Edge, a unified latent-space encryption framework that couples a lightweight image encoder (LiE-Net) with a novel two-pass symmetric cipher called Block-Spiral XOR (BSX) and an outer LDPC channel code. LiE-Net compresses the input image into a 64-dimensional latent vector accompanied by an external residual stream that preserves high-frequency structure for fidelity-aware reconstruction. The BSX cipher applies a forward keystream chain based on SHA3-256 followed by a spiral feedback pass that propagates the final ciphertext block back to the first, yielding global diffusion at constant memory cost. We characterise LSE-Edge across security, robustness, and efficiency. Across the USC-SIPI, ImageNet-IoT, and a curated edge-camera dataset, BSX achieves NPCR of 99.71%, UACI of 33.58%, and ciphertext entropy of 7.998 bits, exceeding the AES-CBC and prior chaotic, DNA-hybrid, and Vision-Transformer-cipher baselines. With LDPC rate-1/2 protection, the framework retains SSIM above 0.74 at channel noise sigma equal to 20, where uncoded variants collapse to SSIM below 0.15. End-to-end latency is 311 milliseconds on a Jetson Xavier and 488 milliseconds on a Raspberry Pi 4 for a 256x256 RGB image, with the BSX stage itself contributing only nine to twenty-eight milliseconds. The combined results indicate that latent-space encryption with a small, well-designed cipher and an outer error-correcting layer can replace bulk pixel encryption for typical edge-AI vision pipelines without sacrificing security or reconstruction quality.
