Multi-layer machine-learning adversarial threat defense in voice and chat channels

Abstract

Methods and systems, including computer-readable media, are described for detecting fraudulent calls or chat messages based on machine-learning (ML) techniques. A computing platform uses multi-layer ML adversarial threat defense systems to proactively detect deceptive bots that carry out fraudulent calls. The systems can identify data derived from signals corresponding to a communication received from a source. A classifier model computes a probability that the source is a non-human entity based on the data provided as an input to the classifier model. An adversarial ML inference engine is used to generate a reply prompt based on the probability that the source is a hon-human entity being provided as an input to the classifier model. The system converts the reply prompt to a communication signal for output to the source and transmits the communication signal to the source as a reply to the communication received from the source.