Struct PerplexityAnalyzer 

pub struct PerplexityAnalyzer { /* private fields */ }

Perplexity analyzer for detecting adversarial patterns

Uses character-level n-gram models and entropy calculations to detect unusual text patterns that may indicate adversarial suffixes or manipulated content.

Implementations

impl PerplexityAnalyzer

pub fn new() -> Self

Create a new analyzer with default English n-gram model

pub fn with_config(config: PerplexityConfig) -> Self

Create a new analyzer with custom configuration

pub fn char_perplexity(&self, text: &str) -> f32

Calculate character-level perplexity for text

Perplexity measures how “surprising” the text is according to the character n-gram model. Higher values indicate unusual text.

pub fn token_entropy(&self, text: &str) -> f32

Calculate token/character entropy

Low entropy indicates repetitive patterns (e.g., “aaaaaaa”). Normal text has moderate entropy.

pub fn unique_char_ratio(&self, text: &str) -> f32

Calculate unique character ratio

Very low ratio indicates repetitive text. Very high ratio with long text may indicate random characters.

pub fn find_anomalous_segments( &self, text: &str, max_perplexity: f32, min_perplexity: f32, min_entropy: f32, ) -> Vec<AnomalySegment>

Detect anomalous segments in text using sliding window analysis

Returns segments that have unusual perplexity or entropy scores.

pub fn is_suspicious( &self, text: &str, max_perplexity: f32, min_entropy: f32, ) -> bool

Quick check for adversarial patterns

Returns true if the text shows signs of adversarial manipulation.

pub fn analyze_suffix( &self, text: &str, suffix_ratio: f32, max_perplexity: f32, min_entropy: f32, ) -> Option<AnomalySegment>

Analyze the suffix portion of text

GCG and AutoDAN attacks typically add adversarial suffixes. This method focuses on the last portion of the text.

Arguments

• text - The text to analyze
• suffix_ratio - Portion of text to analyze as suffix (0.1 - 0.5)
• max_perplexity - Maximum allowed perplexity
• min_entropy - Minimum allowed entropy

Trait Implementations

impl Default for PerplexityAnalyzer

fn default() -> Self

Returns the “default value” for a type.