Hidden Markov Preference Model
A probabilistic framework for attack prediction based on hidden Markov models. It treats attacker behavior patterns as hidden states, uses a feasibility-mask tree to model action parameters, and supports online prediction and adaptation to pattern switches.
Timeline (last 90 days)
The hidden Markov preference model proposed in the paper treats the attacker's active behavior modes as hidden states, defines the action regularities under each mode on a 'feasibility mask tree' whose nodes correspond to executable actions, and the mask dynamically blocks infeasible action branches according to the current state.
The same likelihood function of this model is used both to compute the predictive distribution of the next action and to update mode beliefs after actions are observed, and it supports online real-time inference over partially observed action prefixes.
The paper's theoretical analysis derives three guarantees: a reach-weighted node-level KL expression, a sequence-level log-loss bound, and a uniform fallback guarantee.
On the CAGE 2 benchmark, the filtered hidden states are worth 0.8 nats per step.
After a controller switch, the model's loss begins to decrease within 3 steps, whereas the recurrent baseline requires 10 steps.
On this benchmark, the paper compares the HMM with recurrent neural network baselines, using per-step predictive log loss as the metric.
The model's practical deployment still needs to take into account computational overhead and the cost of constructing the tree structure.
All Facts (7)
The hidden Markov preference model proposed in the paper treats the attacker's active behavior modes as hidden states, defines the action regularities under each mode on a 'feasibility mask tree' whose nodes correspond to executable actions, and the mask dynamically blocks infeasible action branches according to the current state.
50%UnverifiedThe same likelihood function of this model is used both to compute the predictive distribution of the next action and to update mode beliefs after actions are observed, and it supports online real-time inference over partially observed action prefixes.
50%UnverifiedThe paper's theoretical analysis derives three guarantees: a reach-weighted node-level KL expression, a sequence-level log-loss bound, and a uniform fallback guarantee.
50%UnverifiedOn the CAGE 2 benchmark, the filtered hidden states are worth 0.8 nats per step.
50%UnverifiedAfter a controller switch, the model's loss begins to decrease within 3 steps, whereas the recurrent baseline requires 10 steps.
50%UnverifiedOn this benchmark, the paper compares the HMM with recurrent neural network baselines, using per-step predictive log loss as the metric.
50%UnverifiedThe model's practical deployment still needs to take into account computational overhead and the cost of constructing the tree structure.
50%