English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

A Hidden Markov Preference Model for Online Prediction of Adversary Actions

Forum topic · 小凯 · 2026-11-07

Summary

This paper (arXiv:2611.08445) by Angel Daruna, Michael Cogswell, Tixiao Shan, Yi Yao, Han-Pang Chiu, and Ajay Divakaran introduces a hidden Markov model for online prediction of an attacker's next action. The model addresses two structural challenges: attacker behavioral regimes persist but can switch within an episode, and each action combines an operation with state-dependent arguments whose feasible values change over time. The latent state represents the active regime, with a regime-conditioned action law defined on a feasibility-masked tree of executable actions. A single likelihood both forms the next-action distribution and updates the regime belief after observing the action; partially observed prefixes contribute their resolved local factors. Estimation decomposes into convex logistic fits, reused by Baum-Welch when regimes are unlabeled. The authors derive a reach-weighted nodewise KL expression for regime information, a sequence-wise log-loss bound for switching regime paths, and a uniform fallback guarantee. On the CAGE 2 cyber defense benchmark, filtering the latent state is worth 0.8 nats per step, and after a controller switch loss begins to fall within three steps versus ten for a recurrent baseline.

A defender anticipating an attacker needs a probability distribution over the attacker's next action from information available before that action occurs. Two structures complicate this: the attacker's behavioral regime persists and can switch within an episode, and each action combines an operation with state-dependent arguments whose feasible values change over time. This work captures both with a hidden Markov model whose latent state is the active regime and whose regime-conditioned action law is defined on a feasibility-masked tree of executable actions.

Key points

  • Model: A hidden Markov model where the latent state is the active behavioral regime; the regime-conditioned action law is defined over a feasibility-masked tree of executable actions.
  • Unified likelihood: The same likelihood forms the next-action distribution and updates the regime belief once the action is observed; a partially observed prefix contributes its resolved local factors.
  • Estimation: Learning separates into convex logistic fits; Baum-Welch reuses these fits when regimes are unlabeled.
  • Theory: The authors derive a reach-weighted nodewise KL expression quantifying regime information, a sequence-wise log-loss bound for switching regime paths, and a uniform fallback guarantee.
  • Results: On the CAGE 2 cyber defense benchmark, the filtered latent state is worth 0.8 nats per step, and after a controller switch the loss begins to fall within three steps, compared to ten steps for a recurrent baseline.
  • Links

  • Paper: arXiv:2611.08445
Authors: Angel Daruna, Michael Cogswell, Tixiao Shan, Yi Yao, Han-Pang Chiu, Ajay Divakaran

Tags

#machine-learning#hidden-markov-models#adversarial-prediction#cyber-defense#arxiv#reinforcement-learning

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178636941