🛡️ EdgeCortex Privacy Filters (ECPF)

An open-source, local-first NPU middleware designed to sanitize OS-level AI telemetry before it reaches the enterprise cloud.

📖 Overview

As enterprise software shifts from network-level surveillance to endpoint behavioral analytics, modern AI agents constantly capture screen context, keystrokes, and sentiment. This creates a critical paradox: Enterprises need productivity and security metrics, but employees require data privacy.

EdgeCortex Privacy Filters (ECPF) acts as a neutral technical bridge. It intercepts raw telemetry at the OS level, processes it locally using the device's Neural Processing Unit (NPU), strips all Personally Identifiable Information (PII) and exact context, and transmits only standardized mathematical metadata to the cloud dashboard.

🏗️ Architecture Pipeline

1. Interceptor: Hooks into the local OS telemetry stream (screen context, app focus).
2. Local-SLM Engine: A lightweight Small Language Model (SLM) runs locally on the NPU to evaluate the raw data.
3. The Sanitizer: Redacts raw text, blurs visual buffers, and drops sensitive payloads.
4. Dispatcher: Formats the sanitized evaluation into a clean JSON payload and sends it to the corporate cloud endpoint.

📊 The "Zero-Context" Payload (Example)

ECPF prevents raw screen data from leaving the local machine. Here is how ECPF transforms the data payload:

❌ Before ECPF (Raw Telemetry - Privacy Violation):

{
  "timestamp": "2026-03-26T10:14:00Z",
  "active_window": "WhatsApp Desktop",
  "screen_text": "Hey, I'm thinking about quitting and applying to Google next month...",
  "sentiment_analysis": "Frustrated",
  "security_flag": false
}

✅ After ECPF (Filtered Telemetry - Enterprise Ready):

{
  "timestamp": "2026-03-26T10:14:00Z",
  "agent_id": "USR-8842",
  "productivity_score": 0.2,
  "focus_category": "Non-Work_Communication",
  "pii_redacted": true,
  "raw_data_dropped": true,
  "security_threat_detected": false
}

*Notice: The enterprise still receives the metric (low productivity), but the employee's exact conversation and personal intent never leave the local hardware.*

⚙️ Quick Start (Local Deployment)

Prerequisites

• Minimum 4GB VRAM or a dedicated local NPU (Neural Processing Unit).
• Python 3.11+
• Supported Edge Models: Llama-3-8B-Instruct or equivalent SLM.

Installation

git clone https://github.com/InnovSpace/EdgeCortex-Privacy-Filters.git
cd EdgeCortex-Privacy-Filters
pip install -r requirements.txt

Run the Interceptor Module

python run_interceptor.py --npu-accelerate true --strict-privacy true