Analysis of Machine Learning Algorithms for Anomaly Detection on Edge Devices

Area of Science:

• Computer Science
• Machine Learning
• Internet of Things (IoT)

Background:

• Internet of Things (IoT) devices generate vast amounts of data but have limited resources.
• Traditional analysis relies on central systems or cloud, hindering real-time processing on edge devices.
• Moving machine learning (ML) capabilities to edge devices (ML at the edge) is an emerging trend.

Purpose of the Study:

• To experimentally analyze the processing of large, imbalanced datasets for ML on edge devices.
• To identify optimal training dataset sizes for ML model performance saturation.
• To evaluate data reduction techniques for balancing datasets and assess their impact on ML performance and resource usage.

Main Methods:

• A large imbalanced dataset (DS2OS) was split into training (80%) and test (20%) sets.
• Training data was progressively reduced (Di datasets) to find performance saturation points (F1 score ≥ 0.95).
• Two data reduction methods (DRi and DCi) were applied to create balanced datasets, followed by ML algorithm evaluation on Raspberry Pi 4.

Main Results:

• ML model performance metrics saturated, achieving an F1 score of 0.95 or higher, when using 20% of the original training dataset.
• Three dataset reduction methods (random sampling, undersampling majority class, clustering) showed comparable ML performance.
• Evaluation on Raspberry Pi 4 demonstrated the feasibility of running ML algorithms with reduced datasets on resource-constrained edge devices, considering training times and memory usage.

Conclusions:

• Effective data reduction strategies enable high-performance ML on edge devices with limited resources.
• Optimizing dataset size is crucial for achieving performance saturation and efficient edge ML deployment.
• The study validates the practical implementation of ML at the edge for IoT applications.