Impact Analysis of Erroneous Data on IoT Reliability

The ability to sense the environment is the cornerstone of the Internet of Things (IoT), which is a rapidly expanding paradigm that is altering the way we interact with machines. IoT enables a range of new services to enhance the lives of endusers. One of these services concerns activity recognition within Ambient Assisted Living which can be used to help people live independently at home for longer. Many of these applications can, however, be prone to failure and vulnerable to attack. Extensive research is therefore required to build towards a secure and sustainable IoT. This work examines activity recognition in a smart home environment using three different classifiers on a well-known activity recognition dataset. Fail-dirty and device shut-down data is introduced in the dataset to examine the impact that this erroneous data has on the application. This study found that it was possible to rank the importance of sensors with regards to their influence on classification by observing how these failures impacted the classifiers when compared to the fmeasure produced from the classification of the clean data. This work also found that while representing data in a binary format obtains higher accuracy, it makes the classifier considerably more vulnerable to dirty data. Lastly, this study found that decision tree classifiers have an inherent vulnerability when it comes to handling dirty data, resulting in a 24% reduction in performance versus the clean data, due to the structuring and placement of leaf nodes in the tree.