panduwana/interview-ai-detector

Interview AI Detector

Playground

https://panduwana-interview-ai-detector.hf.space/docs

Dev setup

Requirements:

• —Python 3.10.x (may have problems with higher versions)
• —Accept the agreement for https://huggingface.co/google/gemma-2b

git clone git@gitlab.com:rekrutmen_plus/interview-ai-detector.git
cd interview-ai-detector

With Docker

<details> <summary>Local build/run</summary>

  cp COPY_THEN_EDIT.env .env # then FILL IT OUT
  pip install python-dotenv transformers huggingface_hub
  python download-huggingface-model.py google/gemma-2b
  docker build -t interview-ai-detector --secret id=dotenv,src=.env .
  PORT=8080 docker run -p $PORT:7860 interview-ai-detector

Then open: http://localhost:8080/docs/ </details> <details> <summary>Using Huggingface Spaces</summary>

• —define a secret: HUGGINGFACE_TOKEN in the space's settings
• —git push --force </details>

Without Docker

python --version # ensure 3.10.x
pip install -r requirements.txt
python huggingface-download-model.py google/gemma-2b
python -m nltk.downloader punkt wordnet averaged_perceptron_tagger
unzip ~/nltk_data/corpora/wordnet.zip -d ~/nltk_data/corpora/
PORT=8080 uvicorn prediction:app --host 0.0.0.0 --port $PORT

Then open: http://localhost:8080/docs/

Overview

Interview AI Detector is a machine learning model designed to distinguish between human and AI-generated responses during interviews. The system is composed of two models:

1. 1.ALBERT Model: Processes text features extracted from responses.
2. 2.Logistic Regression Model (LogReg): Utilizes the output from the ALBERT model along with additional behavioral features to make the final prediction.

The model is deployed on Google Vertex AI, with integration managed by a Kafka consumer deployed on Google Compute Engine. Both the model and Kafka consumer utilize FastAPI for API management.

Architecture

ALBERT Model

• —Source: HuggingFace
• —Input: 25 numerical features extracted from the text, including:
• —Part-of-Speech (POS) tags
• —Readability scores
• —Sentiment analysis
• —Perplexity numbers
• —Output: Features used as input for the Logistic Regression model

Logistic Regression Model

• —Input:
• —Output from the ALBERT model
• —4 additional features, including typing behavior metrics such as backspace count and key presses per letter
• —Output: Final prediction indicating whether the response is human or AI-generated

Deployment

• —Model Deployment: Vertex AI
• —Kafka Consumer Deployment: Compute Engine
• —API Framework: FastAPI
• —Training:
• —Epochs: 8
• —Dataset: 2000 data points (1000 human responses, 1000 AI-generated responses)
• —Framework: PyTorch

Usage

API Endpoints

• —POST /predict:
• —Description: Receives a pair of question and answer, along with typing behavior metrics. Runs the prediction pipeline and returns the result.
• —Input:

    {
      "question": "Your question text",
      "answer": "The given answer",
      "backspace_count": 5,
      "letter_click_counts": {"a": 27, "b": 4, "c": 9, "d": 17, "e": 54, "f": 12, "g": 4, "h": 15, "i": 25, "j": 2, "k": 2, "l": 14, "m": 10, "n": 23, "o": 23, "p": 9, "q": 1, "r": 24, "s": 19, "t": 36, "u": 9, "v": 6, "w": 8, "x": 1, "y": 7, "z": 0}
    }

• —Output:

    {
      "predicted_class": "HUMAN" or "AI",
      "main_model_probability": "0.85",
      "secondary_model_probability": "0.75",
      "confidence": "High Confidence" or "Partially Confident" or "Low Confidence"
    }

Limitations

• —The model is not designed for retraining. The current implementation focuses solely on deployment and prediction.

Author

Yakobus Iryanto Prasethio