Reputation Prediction of Anomaly Detection Algorithms for Reliable System

Abstract

Today, sensors and/or anomaly detection algorithms (ADAs) are used to collect data in a wide variety of applications(e.g. Cyber security systems, sensor networks, etc.). Today, every sensor or ADA in its applied system participates in the collection of data throughout the entire system. The data collected from all of the sensors or ADAs are then integrated into one significant conclusion or decision, a process known as data fusion. However, the reliability, or reputation, of a single sensor or ADA may change over time, or may not be known at all. Since this reputation is taken into account when determining the final conclusion post data classification, one must be able to predict their reputations. We propose a new machine learning prediction technique (MLPT) to predict the reputation of each sensor or ADA. This technique is based on the existing 'Decision Tree Certainty Level' technique, or DTCL, which is the creation of many random decision trees (forests) with high certainty levels [Dolev et al. (2009)]. In particular, it was shown that the DTCL enhances the classification capabilities of CARTs (Classification and Regression Trees) [Briman et al. (1984)]. After applying the DTCL technique to the reputation data, we then apply a new evolutionary process on those decision trees to reduce the overall number of trees by merging only the most accurate trees and then using only these new trees to generate the reputation values. Thus, we combine DTCL and evolution techniques to enable the determination of sensor or ADA reputations by using only the most accurate trees. Finally, we demonstrate how to improve the data fusion process by identifying the most reliable portions of the collected data to reach more accurate conclusions.

Fraud Detection using Machine Learning

“…• Clustering: This technique groups transactions based on their similarities, potentially uncovering clusters deviating from normal user behavior, like a sudden surge of transactions from a specific location [4]. • Anomaly detection: [24] Algorithms spotlight data points significantly deviating from established patterns, potentially signaling fraudulent activity [5].…”

“…Theoretical Framework: Guiding Our Pursuit Guiding our research is the "Anomaly Detection [24]" framework, which posits that fraudulent transactions deviate significantly from the normal patterns of legitimate activity. By understanding these patterns and identifying deviations, we can effectively pinpoint potentially fraudulent cases.…”