Show HN: Prompt injection detector beats ProtectAI by 19% accuracy, 8.9x smaller

Category: ai-ml

Tags: llm-security, prompt-injection, onnx-optimization, adversarial-ai

Score: 7.7/10 (Innovation: 7, Technical: 8, Documentation: 7, Utility: 8)

A production-grade prompt injection detection model that outperforms the leading competitor with 19% higher accuracy while being 8.9x smaller through INT8 quantization. It's interesting because it combines multiple SOTA NLP techniques (EDL, DoRA, SupCon, etc.) to achieve high precision (95.84%) while maintaining low CPU latency (~101ms), making it practical for real-time AI firewall applications.

Target audience: backend devs, devops, ai-engineers

Repository: https://huggingface.co/hlyn/prompt-injection-judge-deberta-70m

View on Hacker News