Computational Offloading for CNN-based Toxic Comment Detection on a Smartwatch

Smartwatches are an important enabler of the Social Internet of Things (SIoT). However, a successful transition to SIoT will require negotiating challenges specific to social networks. One current challenge for social networks is the detection and removal of toxic comments like insults, threats, or sexually explicit language. Many proposed techniques for detecting toxic comments use deep neural networks. Like Siri, a smartwatch can use a remote service to detect toxic comments, or alternatively run the neural network on the edge to detect such comments. This paper presents the results of an experiment comparing the tradeoffs in memory consumption, CPU load and response time between running a toxic text detection CNN on a Samsung S3 smartwatch, or running the CNN remotely using computational offloading. Sentences were processed either periodically or by using a Poisson distribution with periods of between 0.25 and 4 minutes. The results were that there was little difference in battery depletion between running the CNN locally on the watch or remotely running the CNN. However, using WIFI for offloading resulted in much better (< 1 second) response time than running the CNN on the watch (1–2 seconds). This suggests that computational offloading is a preferred solution in this instance.