Detection of large scale influence operations using machine learning techniques

Abstract

With the ever-increasing popularity of social media, and the rising prevalence of fifth generation warfare that relies heavily on non-kinetic techniques such as propaganda, social engineering, hacking, etc. There has been a significant increase of large scale influence operation, which has become significantly easier to carry out with the recent advance- ments in generative AI and large language models, this has increased prejudice between communities, which has in turn decreased the tolerance seen between each other. In this dissertation we propose a technique which allows for the enumeration of both explicit and implicit biases found in one or multiple communities using a technique known as WEAT (Word embedding association tests), we mainly use it to find problematic associations made by these communities and how these associations change over time in response to outside influence. Using our technique, we were able to achieve an average P value of 0.033 and have been able to show clear problematic associations made by different communi- ties, we also detected sudden shifts in associations which correlated with related outside events