NLP_Adv_Prompt_Detector

Fine‑tuning DistilBERT and RoBERTa as a hierarchical classifier to intercept and block adversarial prompts before they reach any LLM.

Adapting DistilBERT and RoBERTa into a multi‑level filtering system that prevents adversarial inputs from ever reaching a large language model.

Leveraging fine‑tuned DistilBERT and RoBERTa in a tiered classifier architecture to detect and stop adversarial prompts from passing through to any LLM.

Here we used GPT-4. Any LLM can be used in a similar way to safegaurd it from JailBreak Attemps. Plus we can always fine-tune the classifiers accordingly.

Here we used AWS Sagemaker for fine-tuning both DistilBERT and RoBERTa models.

Acknowledgment

This project was developed as part of a collaborative academic effort.

Contributors

This project was developed by:

• Abenanth Gurunathan
• Bertrand Unwan