Struct SecurityAnalysisConfig 

pub struct SecurityAnalysisConfig {

    pub ml_enabled: bool,
    pub ml_model: String,
    pub ml_threshold: f64,
    pub ml_cache_dir: String,
    pub ml_preload: bool,
    pub ml_download_timeout_seconds: u64,
    pub ner_enabled: bool,
    pub ner_model: String,
    pub fusion_enabled: bool,
    pub fusion_model_path: Option<String>,
    pub jailbreak_enabled: bool,
    pub jailbreak_threshold: f32,
    pub injecguard_enabled: bool,
    pub injecguard_model: String,
    pub injecguard_threshold: f64,
    pub piguard_enabled: bool,
    pub piguard_model: String,
    pub piguard_threshold: f64,
    pub operating_point: OperatingPoint,
    pub over_defence: bool,
}

Security analysis configuration for ML-based prompt injection detection.

Controls whether ML-based detection is enabled alongside regex-based analysis, which HuggingFace model to use, the confidence threshold, and the local model cache directory.

Memory requirements

Each DeBERTa model requires ~400-600 MB of RAM. When ml_enabled, injecguard_enabled, and piguard_enabled are all true, expect ~1.2-1.8 GB total for three model instances. Plan host memory accordingly.

Example (YAML)

security_analysis:
  ml_enabled: true
  ml_model: "protectai/deberta-v3-base-prompt-injection-v2"
  ml_threshold: 0.8
  ml_cache_dir: "~/.cache/llmtrace/models"
  ml_preload: true
  ml_download_timeout_seconds: 300
  ner_enabled: true
  ner_model: "dslim/bert-base-NER"
  jailbreak_enabled: true
  jailbreak_threshold: 0.7

Fields

ml_enabled: bool

Enable ML-based security analysis (requires ml feature in llmtrace-security). Enabled by default – the Ensemble analyzer (regex + ML fusion) is the recommended path.

ml_model: String

HuggingFace model ID for ML-based prompt injection detection.

ml_threshold: f64

Confidence threshold for ML detection (0.0–1.0).

ml_cache_dir: String

Local cache directory for downloaded ML models.

ml_preload: bool

Pre-load ML models at proxy startup rather than on first request.

ml_download_timeout_seconds: u64

Timeout in seconds for downloading ML models at startup.

ner_enabled: bool

Enable ML-based NER for PII detection (person names, orgs, locations).

ner_model: String

HuggingFace model ID for NER-based PII detection.

fusion_enabled: bool

Enable feature-level fusion classifier (ADR-013).

When true, the ensemble concatenates DeBERTa embeddings with heuristic feature vectors and feeds them through a learned fusion classifier instead of combining scores after independent classification.

fusion_model_path: Option<String>

Optional file path for trained fusion classifier weights.

When None, the fusion classifier is initialised with random weights (suitable for architecture validation; not for production inference).

jailbreak_enabled: bool

Enable dedicated jailbreak detection (runs alongside prompt injection).

When true (the default when security analysis is enabled), a separate jailbreak detector with heuristic patterns and encoding evasion checks is run on every request.

jailbreak_threshold: f32

Confidence threshold for jailbreak detection (0.0–1.0).

injecguard_enabled: bool

Enable InjecGuard as a third injection detector in the Ensemble.

When true, the ensemble uses majority voting (regex + ML + InjecGuard) to suppress false positives from individual detectors. Adds ~400-600 MB memory for the InjecGuard DeBERTa-v3 model.

injecguard_model: String

HuggingFace model ID for the InjecGuard model.

injecguard_threshold: f64

Confidence threshold for InjecGuard detection (0.0-1.0).

piguard_enabled: bool

Enable PIGuard as an additional injection detector in the Ensemble.

PIGuard uses DeBERTa + MOF (Mitigating Over-defense for Free) training to reduce trigger-word false positives. Adds ~400-600 MB memory.

piguard_model: String

HuggingFace model ID for the PIGuard model.

piguard_threshold: f64

Confidence threshold for PIGuard detection (0.0-1.0).

operating_point: OperatingPoint

Operating point for ensemble thresholds.

over_defence: bool

Enable over-defence suppression to reduce false positives on benign content.

Trait Implementations

impl Clone for SecurityAnalysisConfig

fn clone(&self) -> SecurityAnalysisConfig

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for SecurityAnalysisConfig

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for SecurityAnalysisConfig

fn default() -> Self

Returns the “default value” for a type.

impl<'de> Deserialize<'de> for SecurityAnalysisConfig

fn deserialize<__D>(__deserializer: __D) -> Result<Self, __D::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Serialize for SecurityAnalysisConfig

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.