Background

Fall detection has been extensively explored in both academia and industry. Different methods with different levels of accuracy have been developed for detecting people falling. Using cameras and image processing [1], location and depth sensors [2], and wearable sensors which is the main focus of the current work. Authors of SmartFall [3] used the data collected from a microsoft band-2 smartwatch to train a deep recurrent neural network for detecting people falling. Up-Fall [4] takes a multimodal approach by combining fall data from wearable sensors, visual sensors, and ambient sensors. Multiple useful datasets are included in this dataset. They developed a large multimodal dataset [5] for fall detection. The survey in [6] is similar to [4] and has listed multiple useful fall datasets. There are also multiple publicly available github repos of completed projects for fall detection [7-8].

Adversarial machine learning has been extensively researched in the past few years and the surge of research and attention in deep neural networks. This survey [9] provides a thorough introduction of different adversarial attack scenarios and their terminology. Using adversarial machine learning for training a neural network to detect falling in an unsupervised manner [10], while not considered security-related research, provides useful insight into the usage of these systems for fall detection. Different adversarial attack scenarios will be tested on fall detection systems during different levels.