Systems and Methods of Using Language Models for Combating Online Fraud

A method of detecting fraudulent messages using a server computer system connected to a plurality of client devices, the method comprising employing at least one hardware processor of the server computer system to:

• receive a compact message embedding from a client device of the plurality of client devices, the compact message embedding characterizing a natural language (NL) message received at the client device, the compact message embedding computed by a client-side language model (LM) pre-trained to identify fraudulent messages;
• inflate the compact message embedding to produce an inflated embedding substantially larger than the compact message embedding;
• input the inflated embedding to a selected component of a server-side LM executing on the server computer system, the server-side LM substantially larger than the client-side LM;
• and
• determine whether the NL message is fraudulent according to an output of the server-side LM.

A non-transitory computer-readable medium storing instructions which, when executed by at least one hardware processor of a server computer system connected to a plurality of client devices, cause the computer system to:

• receive a compact message embedding from a client device of the plurality of client devices, the compact message embedding characterizing a natural language (NL) message received at the client device, the compact message embedding computed by a client-side language model (LM) pre-trained to identify fraudulent messages;
• inflate the compact message embedding to produce an inflated embedding substantially larger than the compact message embedding;
• input the inflated embedding to a selected component of a server-side LM executing on the server computer system, the server-side LM substantially larger than the client-side LM;
• and
• determine whether the NL message is fraudulent according to an output of the server-side LM.