Forecasting Cyber Threat Intelligence with Memory Augmented Transformer

Abstract

Cyber threat intelligence data are volatile, irregular, and shaped by abrupt regime shifts, making accurate forecasting particularly challenging. Motivated by this, we explore the potential of a memory-augmented Transformer forecaster that integrates an evolving memory mechanism and confidence-regulated attention. Introducing complementary design that enables the model to balance adaptability with stability, remaining robust under noise and structural changes in the threat landscape. Building on and re-architecting the original ACWA-based approach, the resulting ChronoTensor introduced enhanced model achieves parity with state-of-the-art forecasting methods while introducing transparent memory and attention pathways that enhance the interpretability and explainability of its predictions.