Abstract
Mobile Edge Crowdsensing (MECS) faces a critical trade-off between data utility and privacy protection under dynamic environmental risks and device resource constraints. We present ALPINE, a lightweight, closed-loop framework that enables resource-constrained IoT devices to adaptively adjust Differential Privacy (DP) levels in real-time. By integrating a scalable autoencoder (LightAE) for risk perception and a Twin Delayed Deep Deterministic Policy Gradient (TD3) reinforcement learning agent for privacy budgeting, ALPINE optimizes the balance between privacy gain, data utility, and energy consumption. Experimental results on Raspberry Pi and Arduino platforms demonstrate high anomaly detection accuracy and effective privacy protection with minimal system overhead.
Introduction
Mobile Edge Crowdsensing (MECS) relies on distributed IoT devices, such as wearables and smart city sensors, to collect and transmit data to edge servers. However, data transmission over wireless channels is highly susceptible to eavesdropping and inference attacks. While Differential Privacy (DP) is a standard defense, traditional implementations use a static privacy budget ($\epsilon$). In dynamic real-world environments, static DP is inefficient: it may inject excessive noise when risks are low, destroying data utility, or insufficient noise when risks are high, leading to data leakage. Furthermore, the computational cost of adaptive privacy models often exceeds the capabilities of battery-powered edge devices.
The ALPINE Framework
ALPINE is an end-to-end framework designed to shift heavy model training to an offline phase, enabling lightweight, real-time inference on edge devices. The system operates as a closed-loop architecture comprising four primary modules:
- Dynamic Risk Perception: Calculates a multi-dimensional risk score based on channel stability, data semantics, and resource usage. This is powered by LightAE, a novel scalable autoencoder.
- Privacy Decision: A TD3 reinforcement learning agent determines the optimal privacy budget ($\epsilon$) by balancing privacy gain, data utility loss, and energy costs.
- Privacy Execution: Implements a Bounded Laplace (BLP) mechanism to inject noise while keeping perturbed data within physically plausible bounds.
- Performance Verification: The edge server evaluates data utility and privacy resilience, providing feedback to the terminal device to refine the TD3 agent's policy.
Lightweight Risk Perception with LightAE
A core innovation of ALPINE is LightAE, a block-granularity scalable autoencoder designed for resource-constrained environments. Unlike static models, LightAE can dynamically swap neural network blocks on the fly to adapt to current CPU and memory availability. This allows the device to maintain high anomaly detection accuracy without exceeding real-time latency or memory constraints during resource spikes.
Experimental Evaluation
The authors evaluated ALPINE using diverse datasets, including custom network monitoring data (FD, SD) and public datasets like KDD CUP HTTP and the Diabetes 130-US Hospitals dataset.
Key Results
- Efficiency: LightAE achieved anomaly detection F1 scores between 95.4% and 96.2% while consuming less than 2MB of memory.
- Privacy Protection: ALPINE successfully suppressed Membership Inference Attacks (MIA) to an AUC of ~0.5 (random guess) under high-risk scenarios.
- System Overhead: Deployment on a Raspberry Pi 5 added only ~121ms of latency and 0.32W of energy consumption. On an Arduino Portenta H7, it added just 80ms of latency and ~2.5% memory overhead.
Conclusion and Future Work
ALPINE demonstrates that adaptive, RL-based privacy policies are viable for the edge when combined with scalable architectures like LightAE. The closed-loop feedback mechanism ensures that privacy budgets remain tuned to actual downstream utility needs. Future work will focus on optimizing server-side throughput for massive device populations and exploring even lighter anomaly detection models to further reduce terminal overhead.
Summary updated