Federated Multi-Modal AI for Insider Threat Prediction in Hybrid Workforce Environments

Abstract

The sudden move to hybrid working – where employees split their time between the office and home – is just making the challenge of insider threat detection more complex. Centralized machine learning is limited by data privacy and multi-modal integration, as well as its ability to adapt with distributed endpoints. This paper presents a new framework known as Federated Multi-Modal Artificial Intelligence (FMM-AI) for predicting insider threats, which fuses behavioural, textual, network and physiological modalities from different organisations while not exchanging raw data. Based on federated learning (FL), it supports cross-domain model training, data locality and regulatory requirements. Multi-modal fusion schemes that combine deep neural encoders learnt on each modality and attention-based fusion layers are able to capture such contextual information across modalities. The approach uses Graph Neural Networks (GNN) and Temporal Convolutional networks (TCNs) for detecting subtle behavioral anomalies that may expose insider risks. Results on synthetic hybrid workforce datasets show an increase in predictive accuracy, early detection latency and interpretability w.r.t. competing standalone centralised models. The results demonstrate the promise of FMM-AI to achieve a balance between privacy retention, overhead reduction, and real-time adjustability for providing an innovative means of protecting distributed enterprise environments from insider threats during this era of hybrid work.