Detecting Adversarial Fake Tasks in Mobile Crowd sensing Platforms

The presence of fake tasks in mobile crowd sensing severely affects the normal operation of the platform. Due to the rapid development of deep learning, “Adversarial Fake Tasks” now have a greater destructive impact than traditional fake tasks. Traditional fake tasks are generated based on experience, while adversarial fake tasks are created using deep learning techniques to mimic real tasks, making them difficult to be detected. This paper proposes a “Spatial attention mechanism guided by distance” based on the multi-head self-attention mechanism to counter the threat of adversarial fake tasks in mobile crowdsensing platforms. This module captures the distance correlation between features of generated fake tasks and designs a detection model called Self Distance-Attention Convolution Crowdsensing Network (SDACC). Using SDACC as a detector, the detection steps based on the adversarial training defense strategy are as follows: firstly, utilize the generation model-SeqGAN to learn the distribution of real tasks and generate adversarial fake tasks; then train SDACC to detect adversarial fake tasks. Experiments on the MCS task datasets, the KDD Cup99 dataset, and the CICIDS-2017 dataset demonstrate that SDACC outperforms traditional detection algorithms in the MCS fake task detection domain, with average ACC, F1, and AUC reaching 0.9357,0.9332, and 0.9786 respectively.