Abstract: An email attempting to perpetrate a business email compromise (BEC) attack is detected based on similarity of the email to a known BEC email and on similarity of the email to a user email that would have been sent by the purported sender of the email. Metadata of the email is extracted and input to a BEC machine learning model to find the known BEC email among BEC email samples. The extracted metadata are also input to a personal user machine learning model of the purported sender to generate the user email.

Abstract: Targeted email attacks are detected using feature combinations of known abnormal emails, interflow shapes formed by an email with other emails, or both. An email received in an endpoint computer system is scanned to identify abnormal features indicative of a targeted email attack and the abnormal features of the email are checked against abnormal feature combinations. The email can also be scanned to identify an interflow shape formed by the email with other emails and the interflow shape is checked against interflow shapes of known targeted email attacks.