DRMD: Deep Reinforcement Learning for Malware Detection under Concept Drift

The paper introduces a novel framework called Slot Transformer Imagination with Causality-aware reinforcement learning (STICA) aimed at enhancing deep reinforcement learning agents' efficiency. Traditional world models struggle with complex environments characterized by high-dimensionality and rich object interactions. STICA addresses this by representing observations as object-centric tokens, allowing for better prediction of dynamics and decision-making, akin to human perception of environments.

The study explores data-driven learning of robust stochastic control for infinite-horizon systems with continuous state and action spaces. It highlights the fragility of learned policies in traditional Markov control models due to internal dependencies and external perturbations. The authors propose a distributionally robust stochastic control paradigm that enhances policy reliability by introducing adaptive adversarial perturbations while maintaining the tractability of the Markovian framework.

The article presents RETROFIT, a novel continual learning method designed for security applications. Traditional deep learning models often struggle to adapt to evolving threat landscapes, leading to performance degradation. RETROFIT addresses this by enabling effective knowledge transfer without the need for historical data, thus mitigating the challenges of forgetting while integrating new information.