Struct PatternModel 

pub struct PatternModel<T: Float + Debug + Send + Sync + 'static> { /* private fields */ }

Pattern recognition model for detecting malicious behaviour.

Prototypes are unit-norm gradient directions learned online: honest updates accepted by a round feed normal_patterns, updates rejected as Byzantine feed attack_patterns.

Implementations

impl<T: Float + Debug + Send + Sync + 'static> PatternModel<T>

pub fn new() -> Self

Create new pattern model

pub fn matching_threshold(&self) -> f64

Cosine similarity above which an observation is merged into an existing prototype instead of creating a new one.

pub fn pattern_counts(&self) -> (usize, usize)

Number of learned normal and attack prototypes.

pub fn learn_normal(&mut self, gradient: &Array1<T>) -> Result<()>

Learn a normal behaviour prototype.

pub fn learn_attack(&mut self, gradient: &Array1<T>) -> Result<()>

Learn an attack prototype.

pub fn compute_pattern_deviation( &self, gradient: &Array1<T>, ) -> Result<Option<f64>>

Pattern deviation score in [0, 1], or None when nothing comparable has been learned yet.

The score is the larger of the novelty relative to the closest normal prototype and the alignment with the closest known attack prototype. None is returned whenever no prototype is dimension-compatible with gradient (including after a model reshape, when prototypes exist but none of them applies), so the caller never folds a constant into a combined score.

pub fn compute_pattern_distance( &self, gradient: &Array1<T>, pattern: &Array1<T>, ) -> Result<T>

Euclidean distance between a gradient and a prototype.

Trait Implementations

impl<T: Float + Debug + Send + Sync + 'static> Default for PatternModel<T>

fn default() -> Self

Returns the “default value” for a type.