Combining attack trees and fuzzy numbers in a multi-agent approach to fraud detection