Abstract: Techniques are disclosed relating to methods that include training, by a new profile process executing on a computer system, a contrastive loss function to identify fraudulent images associated with a particular entity. The new profile process may receive new profile information that includes a new profile image and a new profile identifier and compare the new profile identifier to one or more existing profile identifiers. In response to determining that one or more existing profile identifiers satisfy a threshold identifier metric, a particular neural network, using the contrastive loss function, may compare the new profile image to one or more existing profile images corresponding to the one or more existing profile identifiers. The new profile process may determine, using the comparing, whether the new profile information is a possible fake profile of a legitimate profile.