| Abstract: |
The rapid proliferation of edge computing devices equipped with artificial intelligence capabilities has created a compelling demand for image processing pipelines that simultaneously satisfy stringent energy budgets and rigorous data-privacy constraints. This paper presents an empirical investigation of Energy-Efficient and Privacy-Preserving (E2P) image processing for Edge AI systems, introducing the E2P-EdgeNet framework that integrates structured model compression, adaptive quantization, and a novel Privacy-Preserving Differential Privacy (PPDP) module. Experiments conducted across seven benchmark datasets including CIFAR-10, ImageNet, MS-COCO, and a medical chest X-ray corpus demonstrate that E2P-EdgeNet achieves an average energy reduction of 52.7% relative to full-precision baselines while incurring an accuracy degradation of only 0.8 percentage points. Privacy evaluations using differential-privacy budget (epsilon = 1.2) and Peak Signal-to-Noise Ratio (PSNR = 31.8 dB) confirm that the proposed PPDP module preserves data confidentiality without imposing prohibitive inference latency. Deployment benchmarks on three heterogeneous edge platforms Raspberry Pi 4, NVIDIA Jetson Nano, and Google Coral Edge TPU corroborate the framework's cross-platform viability, yielding latency reductions of up to 43 milliseconds and power consumption as low as 540 mW. Comparative analysis against six recent state-of-the-art methods confirms that E2P-EdgeNet consistently outperforms existing approaches on the energy–privacy–accuracy trade-off curve, establishing a new benchmark for resource-constrained Edge AI image processing. |
