System and method for email account takeover detection and remediation utilizing anonymized datasets
A new approach is proposed to support account takeover (ATO) detection based on login attempts by users. The approach relies on assessing fraudulence confidence level of login IP addresses to classify the login attempts by the users. A plurality of attributes/features in one or more user login data logs are extracted and used to build a labeled dataset for training a machine learning (ML) model that relies on statistics of the login attempts to classify and detect fraudulent logins. These attributes make it possible to ascertain if a login attempt or instance by a user is suspicious based on the ML model. In some embodiments, the ML model is trained using anonymized user login data to preserve privacy of the users and a proper level of data anonymization is determined based on the ML model's accuracy in detecting the ATO attacks when trained with different versions of the anonymized data.
1. A system to support account takeover (ATO) attack detection and monitoring, comprising:
an artificial intelligence (AI) engine running on a host, which in operation, is configured to
collect data of a plurality of login attempts to a plurality of user accounts on an electronic messaging system;
generate a labeled dataset based on reputations of IP addresses of a set of random samples of the plurality of login attempts with known good or bad reputation of their IP addresses;
anonymize the labeled dataset of the plurality of login attempts to preserve privacy of users of the plurality of login attempts;
utilize the anonymized labeled dataset of the plurality of login attempts to train a machine learning (ML) model for detection of an ATO attack, wherein the anonymized labeled dataset of the plurality of login attempts is processed to yield an augmented dataset that provides statistical data of the plurality of login attempts for training of the ML model;
detect and classify a new login attempt by the trained ML model to determine if the new login attempt is fraudulent or not.
2. The system of claim 1 , wherein:
the electronic messaging system is one of Office365/Outlook, Slack, LinkedIn, Facebook, Gmail, Skype, Salesforce, and any communication platform configured to send and/or receive the plurality of login attempts to the plurality of user accounts on the electronic messaging system.
3. The system of claim 1 , wherein:
each of the users is either a person or a system or a component configured to login to the electronic messaging system.
4. The system of claim 1 , wherein:
the AI engine is configured to apply different variants of data anonymization to anonymize the labeled dataset of the login attempts.
5. The system of claim 4 , wherein:
the AI engine is configured to apply an IP-subnet to hide a portion of the full IP address of each of the login attempts in the labeled dataset.
6. The system of claim 4 , wherein:
the AI engine is configured to adopt a country attribute to omit and replace the IP address from the labeled dataset.
7. The system of claim 4 , wherein:
the AI engine is configured to adopt a country attribute to omit and replace the IP address from the labeled dataset.
8. The system of claim 4 , wherein:
the AI engine is configured to suppress personally identifying information (PII) portion including names and emails of the users from the labeled dataset.
9. The system of claim 4 , wherein:
the AI engine is configured to compute accuracy and/or precision for the different variants of data anonymization to determine drop in the ML model's accuracy in detecting fraudulent login attempts as a result of the dataset anonymization.
10. The system of claim 4 , wherein:
the AI engine is configured to determine a proper level of data anonymization based on the ML model's accuracy when trained with different variants of data anonymization.
11. The system of claim 1 , wherein:
the AI engine is configured to determine an optimal set of parameters of the labeled dataset for training of the ML model over different combinations of the parameters.
12. The system of claim 1 , wherein:
the AI engine is configured to
combine the IP address of the new login attempt with a set of features extracted and analyzed from the collected data of the plurality of login attempts;
feed the combined IP address of the new login attempt and the extracted set features into the trained ML model to classify the new login attempt;
output determination of the new login attempt.
13. A computer-implemented method to support account takeover (ATO) attack detection and monitoring, comprising:
collecting data of a plurality of login attempts to a plurality of user accounts on an electronic messaging system;
generating a labeled dataset based on the reputations of IP addresses of a set of random samples of the plurality of login attempts with known good or bad reputation of their IP addresses;
anonymizing the labeled dataset of the plurality of login attempts to preserve privacy of users of the plurality of login attempts;
using the anonymized labeled dataset of the plurality of login attempts to train a machine learning (ML) model for detection of an ATO attack, wherein the anonymized labeled dataset of the plurality of login attempts is processed to yield an augmented dataset that provides statistical data of the plurality of login attempts for training of the ML model;
detecting and classifying a new login attempt by the trained ML model to determine if the new login attempt is fraudulent or not.
14. The computer-implemented method of claim 13 , further comprising:
applying different variants of data anonymization to anonymize the labeled dataset of the login attempts.
15. The computer-implemented method of claim 14 , further comprising:
applying an IP-subnet to hide a portion of the full IP address of each of the login attempts in the labeled dataset.
16. The computer-implemented method of claim 14 , further comprising:
adopting a country attribute to omit and replace the IP address from the labeled dataset.
17. The computer-implemented method of claim 14 , further comprising:
suppressing personally identifying information (PII) portion including names and emails of the users from the labeled dataset.
18. The computer-implemented method of claim 14 , further comprising:
computing accuracy and/or precision for the different variants of data anonymization to determine drop in the ML model's accuracy in detecting fraudulent login attempts as a result of the dataset anonymization.
19. The computer-implemented method of claim 14 , further comprising:
determining a proper level of data anonymization based on the ML model's accuracy when trained with different variants of data anonymization.
20. The computer-implemented method of claim 13 , further comprising:
determining an optimal set of parameters of the labeled dataset for training of the ML model over different combinations of the parameters.
21. The computer-implemented method of claim 13 , further comprising:
combining the IP address of the new login attempt with a set of features extracted and analyzed from the collected data of the plurality of login attempts;
feeding the combined IP address of the new login attempt and the extracted set features into the trained ML model to classify the new login attempt;
outputting determination of the new login attempt.