Social bot metrics

In this paper, we propose metrics for malicious bots that provide a qualitative estimation of a bot type involved in the attack in social media: price, bot-trader type, normalized bot quality, speed, survival rate, and several variations of Trust metric. The proposed concept is that after one detects bots, they can measure bot metrics that help to understand the types of bots involved in the attack and estimate the attack. For that, it is possible to retrain existing bot-detection solutions with a metric label so that the machine-learning model can estimate bot parameters. For that, we propose two techniques for metrics labelling: purchase technique—labelling while purchasing an attack, and Trust measurement technique—labelling during the Turing test when humans try to guess which accounts are bots and which are not. In the paper, we describe metrics calculations, correlation analysis, and an example of a neural network which can predict bots’ properties. The proposed metrics can become a basis for developing social media attack detection and risk analysis systems, for exploring bot evolution phenomena, and for evaluating bot-detectors efficiency dependence on bot parameters. We have also opened access to the data, including bot offers, identifiers, and metrics, extracted during the experiments with the Russian VKontakte social network.