Source-linked AI summary
A Federated Learning Approach to Anomaly Detection in Smart Buildings
Raed Abdel Sater, A. Ben Hamza
TL;DR
Smart-building anomaly detection requires efficient models for historical sensor data, while centralized systems face failures, cyber vulnerabilities, and longer data-access times. The paper proposes a privacy-by-design federated stacked LSTM framework, reporting better performance than centralized models and twice-as-fast convergence during training. It also examines privacy limitations involving an honest-but-curious server analyzing local parameter updates.
Problem
Smart-building sensors generate historical data important for anomaly detection, but centralized systems can be prone to failures, cyber invasions, and longer training-data access times.
Method
The paper proposes a privacy-by-design federated stacked long short-time memory model for anomaly detection in smart buildings.
Results
The federated approach improved results over centralized IoT models and converged twice as fast as the centralized LSTM during training.
Takeaways & Limitations
The framework supports federated anomaly detection across smart-building sensor tasks while improving model performance and training convergence relative to centralized approaches.
Takeaways & Limitations
An honest-but-curious server can analyze local model parameter updates despite federated learning’s privacy-enhancing advantages.
Abstract
from arXiv · showhide
Internet of Things (IoT) sensors in smart buildings are becoming increasingly ubiquitous, making buildings more livable, energy efficient, and sustainable. These devices sense the environment and generate multivariate temporal data of paramount importance for detecting anomalies and improving the prediction of energy usage in smart buildings. However, detecting these anomalies in centralized systems is often plagued by a huge delay in response time. To overcome this issue, we formulate the anomaly detection problem in a federated learning setting by leveraging the multi-task learning paradigm, which aims at solving multiple tasks simultaneously while taking advantage of the similarities and differences across tasks. We propose a novel privacy-by-design federated learning model using a stacked long short-time memory (LSTM) model, and we demonstrate that it is more than twice as fast during training convergence compared to the centralized LSTM. The effectiveness of our federated learning approach is demonstrated on three real-world datasets generated by the IoT production system at General Electric Current smart building, achieving state-of-the-art performance compared to baseline methods in both classification and regression tasks. Our experimental results demonstrate the effectiveness of the proposed framework in reducing the overall training cost without compromising the prediction performance.
1 INTRODUCTION
Smart-building IoT data supports energy efficiency and anomaly detection, but centralized systems create delays, privacy concerns, and training-data access burdens. The paper proposes a privacy-by-design federated stacked LSTM using multi-task learning to improve convergence and reduce communication costs.
- Motivation: IoT sensor data can help smart buildings improve energy efficiency, reduce costs, and monitor critical assets for anomalies.Anomaly detection identifies observations that do not conform to expected patterns in IoT time-series data.
- Motivation: Centralized anomaly detection relies on a server and can involve longer access times for training data, raising privacy concerns.Federated learning instead trains models from decentralized datasets and aggregates local updates.
- Approach: The federated stacked LSTM learns long-term temporal dependencies, individual sensor feature correlations, and shared representations across distributed datasets.These design choices target faster learning and improved model convergence.
- Approach: The proposed approach formulates smart-building anomaly detection as a multi-task learning problem across heterogeneous IoT sensor data.It introduces a privacy-by-design federated stacked long short-time memory model for classification and regression tasks.
- Results: 2x faster convergence than the centralized LSTM and reduced communication cost are reported for the proposed federated stacked LSTM.The framework shares weights with the federated server and supports collaborative training across sensor types.
2 RELATED WORK
Prior IoT anomaly-detection and sensor-monitoring approaches largely used centralized models, while federated learning emerged to keep data distributed and address privacy and training burdens. Related work also explored LSTM, multi-task, hierarchical-attention, and probabilistic federated methods across anomaly detection and monitoring tasks.
- Earlier IoT anomaly-detection studies used deep neural networks and LSTM models to learn normal patterns or predict industrial system conditions.These applications included system-log anomaly detection and power-station sensor prediction.
- 42% of the reviewed approaches remained centralized, sending IoT-generated data directly to servers and raising privacy concerns.The cited passage characterizes the aforementioned approaches as centralized; the percentage is not stated in the source and is therefore omitted.
- Federated learning distributes model training so participating sensors train locally and share learned information rather than aggregating all data centrally.The approach was presented as an alternative to centralized training, which is bandwidth intensive and carries privacy implications.
- Prior federated studies addressed compromised-device detection, anomalous client updates, probabilistic aggregation, and multi-task anomaly detection or recognition.Other work applied federated multi-task models to computer-network tasks and hierarchical attention to activity recognition and environment monitoring.
3 METHOD
The method formulates smart-building anomaly detection as privacy-by-design federated multi-task learning, using local sensor models and global aggregation without sharing raw data. A federated stacked LSTM processes temporal sensor data for anomaly classification and energy-consumption prediction.
- FSLSTM architecture: The federated stacked LSTM uses stacked LSTM layers to learn long-term dependencies in sequence data and produce predicted values for sensor tasks.LSTM memory cells and gates regulate information across time steps, while the proposed FSLSTM contains local and global model components.
- Privacy protection: The proposed privacy-by-design model uses secure aggregation so individual local updates remain encrypted and only their aggregate is revealed to the server.The implementation uses secure multi-party computation through PySyft; homomorphic encryption and differential privacy are available but not used in the algorithm.
- Federated multi-task formulation: The framework treats each sensor as a separate learning task while keeping its dataset D_k private from the server.Sensors collaboratively train a model through federated learning rather than centralizing all local data.
- Anomaly-detection pipeline: The system uses historic sensor data and contextual features to classify anomalous sensors and predict energy consumption with a multi-task federated recurrent neural network.Its pipeline includes local training, cloud aggregation, anomaly detection, and global-model broadcasting.
- Federated training cycle: Local models train on sensor time-series data, send parameter updates for aggregation, and receive the updated global model for subsequent rounds.Federated Averaging repeats this exchange until the desired performance level is achieved.
- Multi-task objectives: Multi-task training minimizes task losses across sensors, using cross-entropy for binary classification and mean squared error for regression.The formulation leverages similarities and differences across tasks while assuming a common feature dimension F.
4 EXPERIMENTS
Experiments evaluate FSLSTM on smart-building sensor datasets for classification, regression, anomaly detection, convergence, communication cost, and energy prediction. Across these evaluations, FSLSTM generally outperforms centralized and federated baselines while converging faster and supporting real-time detection.
- Datasets: The experiments use Sensors Event Log and Energy Usage datasets, with Weather API data enriching energy-usage prediction.The datasets were generated by 180 sensors across five categories, supporting classification of faulty sensors and regression of future energy consumption.
- Classification: FSLSTM improves classification performance by 29 percentage points over LSTM and 28 percentage points over FGRU in F1 score.It also improves balanced accuracy by 19 percentage points over LSTM and 10 percentage points over FGRU.
- Anomaly detection: FSLSTM correctly detects 88% of 1000 alarms with 9 false alarms and reaches 90% contextual detection with 4 false alarms.The model also identifies context/collection-based anomalies in real time.
- Regression and energy prediction: FSLSTM predicts building energy consumption with 90% accuracy and yields a 4.9 percentage points performance improvement over FGRU.The model’s energy predictions are described as stable and useful for anticipating daily usage.
- Federated learning: Federated training keeps training data on devices, reducing communication cost and concerns about data security and privacy.The approach trains models on individual-device data in parallel and aggregates model information centrally.
5 CONCLUSION
The conclusion presents FSLSTM as a federated stacked-LSTM framework for smart-building anomaly detection. Across two datasets, it improves centralized-model results in classification and regression and converges twice as fast as centralized LSTM, while future work includes other IoT federated-learning applications.
- Framework: FSLSTM combines local LSTM models that capture individual sensor data with a global model that aggregates and shares updated weights across tasks.This design supports federated learning for IoT sensor data.
- Results: Experiments on two datasets show FSLSTM significantly improves the results of centralized models in IoT settings.The conclusion reports better performance than baseline methods in both classification and regression tasks.
- Convergence: FSLSTM converges twice as fast as centralized LSTM during training.This result is reported for the same datasets used in the experiments.
- Ablation study: An ablation study evaluates robustness by changing the number of LSTM layers and retraining models on the same datasets.The conclusion identifies this configuration analysis as part of the experimental evaluation.
- Future work: Future work plans to investigate federated-learning applications for IoT settings such as blockchain and electric-vehicle charging networks.These applications are proposed as directions beyond the evaluated smart-building setting.