CroMA: Enhancing Fault-Resilience of Machine Learning-Coupled IoT Applications
Yousef AlShehri, Lakshmish Macheeri Ramaswamy · 2024
Machine learning (ML)-coupled Internet of Things (IoT) applications are becoming increasingly popular in many domains. However, faulty sensor readings pose a significant challenge to IoT-ML applications. Noise is inherent in IoT environments due to issues such as harsh IoT environments and hardware limitations. Unfortunately, the performance of the ML applications quickly degrades when faced with faulty sensor readings. Most current techniques to handle faulty sensor readings require many models, and they attempt to correct faulty data using the entire sample containing both readings of non-faulty and faulty sensors, making them unsuitable for edge ML-IoT applications, where real-time processing and low latency are critical. Unfortunately, these techniques fail, especially when multiple sensors' readings are concurrently faulty (due to simultaneous sensor failures). With the goal of building robust edge IoT-coupled ML applications, this paper proposes CroMA - a proactive approach for overcoming simultaneous failure of sensors. CroMA is unique in that it is based upon a masked autoencoder trained on several randomly masked portions of training samples, corresponding to sensors' readings. This enables the autoencoder to learn to predict sensors' readings through the unmasked portions. CroMA incorporates a novel technique to optimize its prediction accuracy by leveraging sensor correlations-based masking and identically treats all fault types as one type via masking. CroMA's network design includes gated recurrent units (GRUs) to capture long-short-term dependencies and correlations between sensors' readings. We confirm the effectiveness of the CroMA through a series of experiments involving three distinct datasets. The experimental results affirm that CroMA effectively corrects sensor faults, thereby preserving the classification accuracy of the IoT ML-based application in the face of sensor failures.