Large-model-driven multimodal retrieval on mobile and edge devices must balance retrieval accuracy, inference latency, and energy efficiency under heterogeneous hardware constraints. This paper proposes an energy-efficiency-aware multimodal retrieval framework for heterogeneous edge environments. The framework first decomposes the end-to-end model into retrieval-aware functional submodules to preserve cross-modal embedding discriminability under low-precision execution. It then introduces a CPU–NPU–DSP collaborative scheduling mechanism to reduce latency and energy consumption, together with an energy-adaptive control strategy that dynamically adjusts execution paths and inference precision under varying energy budgets. Experiments on MS-COCO image–text retrieval show that the proposed method achieves an mAP of \(68.2\%\) and Recall@5 of \(90.0\%\), while reducing end-to-end latency to 118.5 ms and per-query energy consumption to 145.6 mJ. These results demonstrate an improved balance between retrieval effectiveness and edge-side efficiency.

Efficient storage and fast search with high retrieval accuracy remain key challenges for person image retrieval at large image scales. Low-bit visual representations can reduce storage and computation costs, but existing methods often overlook privacy leakage from residual sensitive information in hash codes, allowing attackers to infer attributes such as identity, gender, or age. To address this problem, we propose a privacy-constrained low-bit representation learning framework for person image retrieval. The framework includes a semantic-consistent low-bit retrieval encoder that preserves ranking structures under extremely low-bit budgets, a privacy-utility decoupling encoder that separates retrieval semantics from privacy information, and a balancing strategy that controls the trade-off between utility and privacy through sample-level weighting and bit-level gating. Experiments show that our method achieves an mAP of \(85.6\%\) and Rank-1 accuracy of \(92.5\%\), while reducing the Attack Success Rate under deep inversion attacks from over \(90\%\) to \(53.2\%\). It also maintains low retrieval latency, requiring only 0.02 ms per query, demonstrating a favorable balance between retrieval utility, privacy protection, and efficiency.

The embedded software may migrate the collected data to the server for DNN computation acceleration, which may compromise privacy. We propose a DNN computation framework that combines TEE and NNA to address the privacy leakage problem. We design an NNA-friendly encryption method that enables NNA to correctly compute the encrypted linear input. Facing the overhead of TEE-NNA interaction, we design a pipeline-based prefetch mechanism that can reduce the TEE interaction overhead. Experimentally, our approach proves to be compatible with a wide range of NPUs and TPUs, and improves the performance by 8-19 times over the TEE scheme.