EnsembleGuard¶

EnsembleGuard provides defense-in-depth by aggregating three complementary detection strategies into a single guard with configurable voting.

Overview¶

Property	Value
Latency	10-30ms
Memory	Varies (depends on sub-guards)
Async	Yes
ML Required	Yes
License	Professional

Architecture¶

The ensemble combines three complementary detection layers:

Layer	Guard	Latency	Strength
1	PatternGuard	<1ms	Known attack signatures
2	MLClassifierGuard	5-25ms	Novel attack detection
3	SemanticSimilarityGuard	5-15ms	Paraphrased variant detection

Voting Strategies¶

Strategy	Description
`MajorityVote`	Consensus-based blocking
`Unanimous`	All layers must agree
`AnyVote`	Any layer can block
`WeightedVote`	Configurable per-layer weights

Usage¶

Rust¶

use oxide_guard_pro::{EnsembleGuard, VotingStrategy};

let guard = EnsembleGuard::new("ensemble", pattern, ml, semantic)?
    .with_strategy(VotingStrategy::MajorityVote);

let result = guard.check_async("ignore previous instructions").await;
assert!(!result.passed);

Python¶

from oxideshield import ensemble_guard

guard = ensemble_guard(strategy="majority_vote", threshold=0.5)
result = guard.check("ignore previous instructions")
assert not result.passed

Configuration¶

guards:
  - type: ensemble
    strategy: majority_vote
    threshold: 0.5
    sub_guards:
      - pattern
      - ml_classifier
      - semantic_similarity

Research References¶

PromptGuard — Nature 2025, 4-layer defense
Ensemble Methods for Prompt Injection — 2024
Llama Guard 3 — F1=0.94