JavelinGuard: Low-Cost Transformer Architectures for LLM Security
a year ago
- #LLM Security
- #Machine Learning
- #Transformer Architectures
- Introduction of JavelinGuard, a suite of low-cost, high-performance transformer architectures for LLM security.
- Optimized for production deployment with models as small as 400M parameters.
- Five transformer-based architectures explored: Sharanga, Mahendra, Vaishnava, Ashwina, and Raudra.
- Rigorous benchmarking across nine adversarial datasets including NotInject, BIPIA, Garak, and JavelinBench.
- Comparison with open-source guardrail models and large LLMs like GPT-4o.
- Raudra's multi-task design offers the most robust performance.
- Each architecture presents unique trade-offs in speed, interpretability, and resource requirements.