| With the use of machine learning growing significantly, it stands to reason that no
industry will be left behind in this field. The nuclear industry is no exception. Although
machine learning and artificial intelligence significantly decrease the time and costs of many
processes and ease the lives of users, the widespread use of machine learning opens the door
to bad actors. This study aims to analyze the cybersecurity risks of a Trojan attack on
nuclear-based neural networks. The viability of these attacks will be examined as well as
potential mitigation techniques that users should implement in the nuclear field. With a
Trojan attack on a nuclear machine learning model that classifies transient and steady state
data, a black hat hacker could change the classifications to fool the model into viewing a
transient as steady state data and vice versa. This poses a significant risk to the nuclear
reactor itself and could have devastating consequences for communities around the nuclear
plant.
Within the scope of this research, a Trojan attack will be developed against a neural
network trained on nuclear datasets, namely the GPWR [ 11] and Asherah datasets [ 17].
After an attack has been successfully implemented and the capabilities of an actor explored,
mitigation techniques will be examined for their effectiveness. Among these techniques are
retraining the model on clean data [ 9], using an autoencoder to detect the Trojaned data
(anomalies) [12], and examining the model weights and parameters to see if a Trojaned model
can be successfully detected [3]. Once the success of each of the techniques above has been
explored, their use cases will be discussed.
Keywords: Trojan attack, Neural Networks, Autoencoder, Machine Learning |