Transformer-based Sentiment Analysis for Anomaly Detection on Drone Forensic Timeline

An IoT device such as a drone is constantly generating log records to store every event that happens to the drone during a flight. In case the drone encounters a problem or experiences an incident, the log can be analyzed to find the root cause. A drone flight log contains a number of parameters, including sensor, state, and message data. These data can be utilized to perform anomaly detection. A common approach to detecting anomalies in log data is measuring the deviation of the log sequence. As an initial attempt, this paper proposes sentiment analysis as an approach for anomaly detection on drone flight log data. We construct our dataset by collecting and annotating the human-readable messages extracted from public datasets. Several existing pre-trained LLMs are fine-tuned to find the best model with the highest evaluation score. The proposed approach can distinguish between anomalous and normal events with 92.527% accuracy.