Botnets Breaking Transformers: Localization of Power Botnet Attacks Against the Distribution Grid
Authors:
Lynn Pepin,
Lizhi Wang,
Jiangwei Wang,
Songyang Han,
Pranav Pishawikar,
Amir Herzberg,
Peng Zhang,
Fei Miao
Abstract:
Traditional botnet attacks leverage large and distributed numbers of compromised internet-connected devices to target and overwhelm other devices with internet packets. With increasing consumer adoption of high-wattage internet-facing "smart devices", a new "power botnet" attack emerges, where such devices are used to target and overwhelm power grid devices with unusual load demand. We introduce a…
▽ More
Traditional botnet attacks leverage large and distributed numbers of compromised internet-connected devices to target and overwhelm other devices with internet packets. With increasing consumer adoption of high-wattage internet-facing "smart devices", a new "power botnet" attack emerges, where such devices are used to target and overwhelm power grid devices with unusual load demand. We introduce a variant of this attack, the power-botnet weardown-attack, which does not intend to cause blackouts or short-term acute instability, but instead forces expensive mechanical components to activate more frequently, necessitating costly replacements / repairs. Specifically, we target the on-load tap-changer (OLTC) transformer, which uses a mechanical switch that responds to change in load demand. In our analysis and simulations, these attacks can halve the lifespan of an OLTC, or in the most extreme cases, reduce it to $2.5\%$ of its original lifespan. Notably, these power botnets are composed of devices not connected to the internal SCADA systems used to control power grids. This represents a new internet-based cyberattack that targets the power grid from the outside. To help the power system to mitigate these types of botnet attacks, we develop attack-localization strategies. We formulate the problem as a supervised machine learning task to locate the source of power botnet attacks. Within a simulated environment, we generate the training and testing dataset to evaluate several machine learning algorithm based localization methods, including SVM, neural network and decision tree. We show that decision-tree based classification successfully identifies power botnet attacks and locates compromised devices with at least $94\%$ improvement of accuracy over a baseline "most-frequent" classifier.
△ Less
Submitted 18 March, 2022;
originally announced March 2022.
