First version uploaded

README.md

@@ -1,3 +1,32 @@
----
-license: mit
----
+# Model Description
+
+The purpose of our trained Random Forest models is to identify malicious prompts given the prompt embeddings derived from [OpenAI](https://huggingface.co/datasets/ahsanayub/malicious-prompts-openai-embeddings), [OctoAI](https://huggingface.co/datasets/ahsanayub/malicious-prompts-octoai-embeddings), and [MiniLM](https://huggingface.co/datasets/ahsanayub/malicious-prompts-minilm-embeddings). The models are trained with 373,120 benign and malicious prompts. We split this dataset into 80% training and 20% test sets. To ensure equal proportion of the malicious and benign labels across splits, we use stratified sampling.
+
+Embeddings consist of fixed-length numerical representations. OpenAI generates an embedding vector consisting of 1,536 floating-point numbers for each prompt. Similarly, the embedding datasets for OctoAI and MiniLM consist of 1,027 and 387 features, respectively.
+
+# Model Evaluation
+
+The binary classification performance of embedding-based random forest models is shared below:
+
+| Embedding | Precision | Recall | F1-score | AUC   | 
+|-----------|-----------|--------|----------|-------|
+| OpenAI    | 0.867     | 0.867  | 0.867    | 0.764 |
+| OctoAI    | 0.849     | 0.853  | 0.851    | 0.731 |
+| MiniLM    | 0.849     | 0.853  | 0.851    | 0.730 |
+
+## How to Use the Model
+
+We have shared three versions of random forest models in this repository. We used the following embedding models: `text-embedding-3-small` from OpenAI, and the open-source models `gte-large` hosted on OctoAI, as well as the well-known `all-MiniLM-L6-v2`. Therefore, you need to covert the prompts to its respective embeddings before querying the model to obtain its  prediction: `0` for benign and `1` for malicous.
+
+
+## Citing This Work
+Our implementation, along with the curated datasets used for evaluation, is available on [GitHub](https://github.com/AhsanAyub/malicious-prompt-detection). Additionaly, if you use our implementation for scientific research, you are highly encouraged to cite [our paper](https://arxiv.org/abs/2410.22284).
+
+```
+@article{ayub2024embedding,
+  title={Embedding-based classifiers can detect prompt injection attacks},
+  author={Ayub, Md Ahsan and Majumdar, Subhabrata},
+  booktitle={CAMLIS},
+  year={2024}
+}
+```

rf_minilm.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5a16a5effa8d82d820dcdc84b4cea9514309dc6423cbd3082949dec78e0e1a43
+size 386568804

rf_octoai.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:11001475d3fe2d296ba84609994b951754627c2798adbc7090e62495014bbb88
+size 358151204