Shatru
Runtime Backdoor Detection Engine for Large Language Models
A black-box runtime backdoor detection engine for large language models built in a 24-hour hackathon sprint. Shatru monitors Shannon Entropy distributions across layers 6–18 of Phi-3-mini to detect Trojan activation patterns on trigger input — completely without access to training data or internal model weights.
System Specifications
key technical metricsSystem Architecture & Methodology
engineering breakdownEntropy Fingerprinting
Monitors statistical divergence in Shannon Entropy across transformer layers 6 through 18 during inference to spot anomalous token distribution spikes.
True Black-Box Defense
Detects backdoors without requiring access to pre-training datasets, weights, or fine-tuning checkpoints — addressing real-world API security.
Supply-Chain Resilience
Provides runtime verification against poisoned open-source models, preventing stealth prompt injection and triggered malicious payloads.
Layerwise Statistical Profiling
Captures baseline activation profiles on clean inputs to compute threshold boundaries, flagging trigger-induced deviations with high statistical sensitivity.