Summary
Current Explainable Artificial Intelligence (XAI) methods for energy management can inadvertently expose sensitive behaviours. In contrast, traditional privacy techniques, such as Differential Privacy (DP), often sacrifice explainability and utility. We propose a Hierarchical XAI (HXAI) pipeline specifically for smart-home energy systems. This pipeline offers households detailed, noise-free explanations (using SHAP, which stands for Shapley Additive Explanations) at the appliance level while providing energy providers with zonal, privacy-preserving summaries generated through Laplace-noised aggregates.
In our HXAI framework, users have control over the extent of information they disclose across multiple requests, something that standard SHAP/LIME approaches and many DP-only methods struggle to manage effectively. The framework guarantees that zonal SHAP remains unbiased in expectation, even with DP noise, thereby maintaining utility while safeguarding individual privacy.
The goal is to build and evaluate a distributed End-Edge (Local/Zonal) explainability pipeline for smart-home energy that prioritizes privacy without compromising the usefulness of the explanations. This involves benchmarking differential privacy (through privacy budgeting with Laplace noise) against non-DP baselines, examining SHAP stability under noise, and quantifying the trade-offs between privacy and utility. The system will ultimately provide localized, household-level insights alongside aggregated reports for energy providers. (Background: HXAI framework.)
Core questions
- How does utility scale with ε for general vs. specific (appliance-level) queries?
- How stable are SHAP attributions after privacy noise at local vs. zonal levels?
- What negotiation rules keep cumulative privacy leakage within bounds across repeated queries?
Tasks
-
Implement a distributed HXAI prototype: local SHAP on households; zonal aggregation with Differential Privacy (DP) with Laplace noise injection; optional privacy databox/TEE stub.
-
Benchmarking: DP vs. non-DP; compare SHAP vs. naive baselines; measure accuracy, explanation drift, and detection of regime changes.
-
Comparative analysis: specific vs. general queries; privacy-utility curves; leakage accumulation across multiple queries.
-
Engineering: data ingestion/cleaning, containerization, reproducible scripts; optional FedAvg for distributed aggregation.
Expected outcomes
-
Reproducible code & docs; benchmark tables/plots; concise thesis & 10-slide defence deck.
Student profile/prerequisites
-
Strong coding (Python: NumPy/Pandas/scikit-learn; PyTorch/TF a plus).
-
Comfortable with basic math/stats (probability, distributions, optimization).
-
Can handle data processing at scale and adapt theory to practice.
Nice-to-have
-
Experience with SHAP/LIME, DP concepts, Docker, FastAPI/Flask, and basic distributed systems.
Datasets
-
Public smart-meter/consumption datasets (or realistic synthetic) + Nordic price series.
Tools/Stack
-
Python, Jupyter, Git, Docker; SHAP; DP utilities; FastAPI (optional); simple DB (SQLite/Postgres).
Deliverables
-
Thesis PDF, code repository, dataset cards, privacy–utility dashboard.
Contacts:
Poushali Sengupta <poushals@ifi.uio.no>
Sabita Maharjan <sabita@ifi@uio.no>