Detecting Anomalies in Simulated Nuclear Data Using Autoencoders

Concerns over cybersecurity in critical systems have grown significantly over the last decade. The increase in the successful attacks against infrastructure, major corporations, and governments has led to major investment in mitigating and preventing cyberattacks. At the same time, there has been a significant interest in utilizing data in operations, with machine learning applications becoming a popular area of study. One industry exploring machine learning applications is the nuclear industry. Because of the sensitive nature of nuclear systems, the question if attacks on nuclear data can be detected has begun to take urgency. This study explores the use of autoencoders to detect anomalies in nuclear data that could be potentially used to evaluate the operating status of a nuclear system. Data from a generic pressurized water reactor simulator used in a previous study to diagnose transients was used to train an autoencoder model using Keras. A separate portion of these data was altered by adding statistical noise for validation. Four different levels of noise were used in this experiment. Once the autoencoder was trained, a threshold was calculated using the average mean square error of the predictions and the standard deviation from that loss. Points above the threshold were classified as anomalies while points below were considered unaltered. For the initial level of noise, the model was able to score near perfect in recall, capturing all but 13 of the 13 884 altered points. However, in terms of precision, the model misclassified a number of unaltered points as altered, resulting in a score of 73.76%. To test the sensitivity of the model, the amount of noise was reduced three times, and as expected, the performance of the model worsened with each reduction. Still, the high performance in identifying altered points for higher levels of noise is an encouraging first step in developing anomaly detection systems for nuclear data.