Abstract
Smart homes generate vast amounts of sensitive data, ranging from health metrics to energy consumption patterns. While fog computing offers a solution for real-time data processing at the network edge, it introduces significant privacy challenges, particularly regarding data sensitivity personalization and resistance to collusion attacks. This paper presents APDP, an Attack-Proof Personalized Differential Privacy model. By integrating a Markov-Laplace mechanism into a four-layer fog computing architecture, APDP successfully decouples data correlations, preventing multiple cloud service providers from neutralizing privacy noise through collusion while maintaining high data utility.
Introduction
The proliferation of IoT devices in smart homes has created a "Big Data" challenge. Traditionally, sensitive information is routed through central gateways to the cloud, but these gateways often struggle with the latency and processing demands of modern smart environments. Fog computing addresses these issues by processing data closer to the source.
However, fog-based systems face two primary hurdles: the lack of personalized privacy settings for varying data types and the risk of collusion attacks. In a multi-cloud environment, providers may share noisy datasets to mathematically "average out" Differential Privacy (DP) noise, exposing the underlying raw data. APDP is designed to mitigate these risks by treating noise as a stochastic process rather than a static additive.
The APDP Framework
The authors propose a robust four-layer architecture consisting of IoT devices, Fog Servers, Public Clouds, and end-user Applications. The core of this system is the APDP model, which rests on three pillars:
- Trust Distance ($D^T$): The system calculates a privacy level ($\epsilon$) based on the network topology. Devices further from the fog server, such as medical wearables, are assigned a higher trust distance and stricter privacy parameters.
- Personalized Differential Privacy (PDP): Unlike global privacy budgets, the fog server assigns unique $\epsilon$ values to different data streams based on their specific sensitivity and intended destination.
- Markov-Laplace Mechanism: To counter collusion, APDP integrates noise generation into a Markov process. By ensuring that noise for one recipient is only related to the preceding state in a stochastic chain, the model breaks the correlations that attackers exploit. Even if multiple clouds combine their data, the noise remains mathematically decoupled.
Experimental Evaluation and Results
The model was evaluated using a real-world health-related smart home dataset covering 15 candidates across seven daily scenarios (e.g., sleeping, hygiene, dressing). APDP was benchmarked against Uniform Differential Privacy (UDP) and Normal Personalized Differential Privacy (NPDP).
Key findings include:
- Collusion Resistance: While standard personalized models (NPDP) eventually fail when multiple clouds collude, APDP remains consistently below the attack threshold, effectively neutralizing the threat.
- Privacy Efficiency: APDP showed a significantly slower increase in the cumulative privacy budget compared to NPDP as more nodes were added to the network.
- Data Utility: APDP achieved a superior trade-off between privacy and utility. As the number of cloud recipients increased, APDP maintained a lower Root-Mean-Square Error (RMSE) than competing models.
Conclusion
APDP represents a significant advancement in smart home privacy, transforming noise generation from a static additive into a dynamic stochastic process. By leveraging the computational power of fog servers to act as a "privacy buffer," the model protects sensitive IoT data against sophisticated collusion attacks without sacrificing the utility required for modern smart applications. Future research will focus on optimizing the model for dynamic environments with mobile IoT devices and utilizing machine learning to automate the privacy-utility trade-off in real-time.
Summary updated