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.