PIIPS - PII Protection Service

📌 Overview

PIIPS (PII Protection Service) is a Flask-based API that provides PII anonymization and de-anonymization using Presidio and integrates with Ollama for processing anonymized data.

Key Features:

• Detects PII (Personally Identifiable Information) in text.
• Anonymizes PII by replacing it with unique placeholders.
• Sends anonymized text to an LLM via Ollama.
• De-anonymizes responses to restore original PII values.
• Provides configurable options (e.g., showing LLM’s internal thought process).

---

📥 Installation & Setup

1️⃣ Clone the Repository

git clone https://github.com/yourusername/PIIPS.git
cd PIIPS

2️⃣ Create a Conda Environment

Ensure you have Conda installed, then create and activate a new environment:

conda create -n piips_env python=3.10 -y
conda activate piips_env

3️⃣ Install Dependencies

pip install -r requirements.txt

4️⃣ Set Up Ollama (Required for LLM Integration)

Ollama provides local execution of LLMs. Install it using:

• Linux/macOS

curl -fsSL https://ollama.com/install.sh | sh

• Windows: Download and install from Ollama's official website.

Then, start the Ollama server:

ollama serve

5️⃣ Download the LLM Model You Want

By default, PIIPS is set to use DeepSeek 7B (deepseek-r1:7b).

To use a any model, download it via Ollama:

ollama pull <model-name>

6️⃣ Specify Your Model in piips.py

Modify piips.py to use your downloaded model. Update this line:

LLM_MODEL = "deepseek-r1:7b"

7️⃣ Run the PIIPS Server

python piips.py

By default, the server runs on port 3000.

🚀 API Endpoints

1️⃣ Health Check

Endpoint: GET /health

Description: Checks if the server is running.

Response:

"PII Protection Server is up"

2️⃣ Anonymize PII

Endpoint: POST /anonymize

Description: Detects and anonymizes PII in a given text.

📤 Request:

{
    "text": "John Doe lives in New York and his phone number is 123-435-6789."
}

📥 Response:

{
    "items": [
        {
            "end": 82,
            "entity_type": "PHONE_NUMBER",
            "operator": "custom",
            "start": 63,
            "text": "<PHONE_NUMBER_8579>"
        },
        {
            "end": 38,
            "entity_type": "LOCATION",
            "operator": "custom",
            "start": 23,
            "text": "<LOCATION_8676>"
        },
        {
            "end": 13,
            "entity_type": "PERSON",
            "operator": "custom",
            "start": 0,
            "text": "<PERSON_6480>"
        }
    ],
    "text": "<PERSON_6480> lives in <LOCATION_8676> and his phone number is <PHONE_NUMBER_8579>."
}

3️⃣ De-anonymize Text

Endpoint: POST /deanonymize

Description: Replaces placeholders in text with original values.

📤 Request:

{
    "text": "<PERSON_5678> lives in <LOCATION_1234>.",
    "deanonymizers": {
        "<PERSON_5678>": "John Doe",
        "<LOCATION_1234>": "New York"
    }
}

📥 Response:

{
    "text": "John Doe lives in New York."
}

4️⃣ Full PII Guard LLM Flow

Endpoint: POST /pii-guard-llm

Description: Anonymizes PII, sends it to an LLM, and de-anonymizes the response.

The show_thinking parameter controls whether the response from the LLM includes its internal reasoning, which appears inside ... tags. Some language models, particularly reasoning models, "think" to themselves before formulating a response. This self-reflection is included in the output as a separate section enclosed in tags.

📤 Request:

{
    "text": "John Doe lives in New York and his phone number is 123-435-6789. Summarize the information provided to you in a table.",
		"show_thinking": false
}

📥 Sample Response (Anonymized & Processed by LLM, then De-Anonymized):

{
    "text": "Here is the summarized information in a table:

    | PERSON      | LOCATION                          | PHONE Number                     |
    |-------------|------------------------------------|-----------------------------------|
    | John Doe | New York                    | 123-435-6789             |
    Note: The phone number placeholder is represented as `123-435-6789` in the table, and the full phone number provided was `123-435-6789`."
}

📝 Notes:

• If "show_thinking": true (default behavior), the response will include the LLM's reasoning (<think>...</think>) if the model outputs such information. Setting "show_thinking": false will remove the LLM's reasoning and only provide the final reponse (as seen in the example above).
• Anonymized text is processed by an LLM (e.g., deepseek-r1:7b) before being de-anonymized back to the original context.

⚙️ Configuration

• The server defaults to port 3000.
• Modify LLM_MODEL in piips.py to use a different LLM.
• By default, GPU acceleration for Ollama is turned on:

  os.environ["OLLAMA_CUDA"] = "1"