Adversarial Machine Learning-Driven Fake Task Anticipation in Mobile Crowdsensing Systems
Zhiyan Chen, Burak Kantarcı · 2021
Mobile Crowdsensing (MCS) enables access to distributed sensing sources of personalized devices so to support various smart services. Services and applications that can benefit from MCS are various such as transportation, health care, public safety, smart mobility and many others. Distributed nature and lack of a pre-established trust mechanism in MCS systems make them vulnerable in the presence of various threats that can be initiated by either sensing data providers or service requesters. While it is relatively easier to detect and eliminate false sensing data submissions via outlier detection, tasks that are submitted to keep the MCS servers and participating devices occupied and clogged are challenging since these attacks can be planned intelligently. In this paper, we investigate the potential of adversarial machine learning to anticipate fake / illegitimate task submissions to MCS systems. To this end, we empower a threat anticipation mechanism that leverages a Generative Adversarial Network (GAN) to inject adversarial samples of fake / illegitimate tasks to an MCS system. We evaluate the impact of GAN-driven attacks in terms of Adversarial Attack Success Rate (AASR) and Attack Severity (AS). Our numerical results show that the potential risk and severity of the offensive use of Machine Learning is significantly higher than an adversarial baseline that injects fake tasks solely based on random noise samples.