Abstract
In the Artificial Intelligence of Things (AIoT), transparency via Explainable AI (XAI) is crucial but introduces significant privacy risks. Explanations, such as those generated by SHAP (SHapley Additive exPlanations), can inadvertently serve as side channels for leaking sensitive user behavior. This paper introduces SHAP entropy regularization, a novel training technique that penalizes models for overly concentrated feature attributions. By encouraging higher entropy in explanations, the proposed SHAP-regularized LSTM maintains high predictive accuracy while effectively mitigating membership inference and fingerprinting attacks.
Introduction
Transparency in smart home environments allows users to understand AI-driven energy forecasts. However, detailed explanations can reveal private routines, such as cooking schedules or specific appliance usage. This creates an "explanation-privacy" paradox: providing meaningful insights without compromising user privacy. Current defenses like Differential Privacy (DP) often incur a heavy "utility tax," significantly degrading model accuracy. This research proposes a "privacy-by-design" approach that regularizes the explanation space itself to protect user data.
Methods
The core of the proposed approach is the integration of privacy constraints directly into the model's learning process.
SHAP Entropy Regularization
Standard models often overfit their reasoning to a few dominant features, creating "spiky" explanations that are easy to fingerprint. The authors introduce a penalty term to the loss function that discourages low-entropy SHAP distributions. This forces the model to distribute feature importance more uniformly across available inputs, effectively "blurring" the reasoning process and making it harder for an attacker to exploit.
SHAP-Regularized LSTM
The technique was implemented using a Long Short-Term Memory (LSTM) network, which is ideal for the sequential nature of energy consumption data. The model was evaluated against a suite of specialized SHAP-based attacks, including entropy attacks, similarity/divergence attacks, and rank correlation attacks designed to identify stable feature importance "fingerprints."
Experimental Setup
The method was validated using the REFIT Electrical Load Measurements dataset, containing real-world power consumption from 20 UK households. The study focused on forecasting aggregate household power based on nine individual appliance monitors (e.g., kettles, dishwashers, TVs) across three specific houses.
Results and Discussion
The SHAP-regularized LSTM was compared against a baseline LSTM and a DP-LSTM.
Quantitative Performance
The regularized model demonstrated a superior balance between privacy and utility. As shown in Figure 1, while the DP-LSTM suffers from increased Mean Absolute Error (MAE), the SHAP-regularized model maintains accuracy close to the baseline. Crucially, it achieves significantly higher SHAP entropy and lower membership similarity, indicating robust protection against inference attacks. For instance, in "House 1," the rank correlation—a measure of fingerprinting ease—dropped from 0.2889 in the baseline to 0.2413 in the regularized model.
Qualitative Visual Evidence
Heatmap analysis reveals the "blurring" effect of the regularization. In the baseline model, specific appliances like the kettle show glaringly obvious attribution spikes during peak hours. In contrast, the regularized model produces more diverse and less "spiky" attributions, making it harder for an adversary to pinpoint specific appliance usage patterns.
Conclusion
SHAP entropy regularization offers a promising path for privacy-preserving XAI in AIoT. By treating explanations as a potential vulnerability and regularizing them during training, models can remain both transparent and secure. Future work will explore adaptive regularization weights and the extension of this method to multimodal AIoT data, while aligning technical metrics with legal frameworks like the GDPR and the EU AI Act.
Summary updated