VFRAME Faceless Web Server

VFRAME Faceless Web Server is a prototype to explore the potential for using face detection/blurring to minimize facial biometrics in web publications. This repo is a prototype for exploring and testing ideas, and feedback is welcome.

Usage

• Setup VFRAME main repo
• Source environment variables for demo script
• Run demo script to generate blurred faces
• Setup apache or nginx config

cd vframe/vframe_cli

# source env vars
source plugins/vframe_faceless_plugin/data/examples/filepaths.sh

# run the CLI process to detect, blur, and save images w/subdirs
# Use ssdface for CPU or yoloface or retinaface models for GPU
./cli.py pipe \
  open -i $DIR_DEMO --recursive \
  detect -m ssdface \
  blur \
  draw \
  save-images -o $DIR_DEMO_OUT --subdirs

Read more about redaction.md

How it works

• Glob all image files in local directories
• Create faceless versions of images
• Save faceless images to replicated directory structure
• Setup Nginx or Apache to serve faceless images for bot-like user agents
• Images are manually processed using CLI scripts, but could be set to a CRON job, or dynamically created using the docker-compose local Flask server API endpoints

Features

• run locally and sync files or remotely using Python CLI
• uses OpenCV DNN CUDA acceleration when available
• includes Model Zoo options for YOLOV4 and RetinaFace
• example Nginx and Apache config scripts for bot-like user agent redirects

Testing

To see the nginx configration in action on this web server, run the following curl commands.

curl 'http://faceless.vframe.io/face.jpg' \
  -H 'Cache-Control: no-cache' \
  -H 'User-Agent: Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_1) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/80.0.3987.132 Safari/537.36' \
  > face.jpg

curl 'http://faceless.vframe.io/face.jpg' \
  -L \
  -H 'Cache-Control: no-cache' \
  -H 'User-Agent: Googlebot/2.1 (+http://www.google.com/bot.html)' \
  > face_removed.jpg

NGINX

• Sample NGINX configuration

Add this user agent detection outside of your server block:

map $http_user_agent $limit_bots {
  default 0;
  ~*(google|bing|yandex|msnbot) 1;
  ~*(AltaVista|Googlebot|Slurp|BlackWidow|Bot|ChinaClaw|Custo|DISCo|Download|Demon|eCatch|EirGrabber|EmailSiphon|EmailWolf|SuperHTTP|Surfbot|WebWhacker) 1;
  ~*(Express|WebPictures|ExtractorPro|EyeNetIE|FlashGet|GetRight|GetWeb!|Go!Zilla|Go-Ahead-Got-It|GrabNet|Grafula|HMView|Go!Zilla|Go-Ahead-Got-It) 1;
  ~*(rafula|HMView|HTTrack|Stripper|Sucker|Indy|InterGET|Ninja|JetCar|Spider|larbin|LeechFTP|Downloader|tool|Navroad|NearSite|NetAnts|tAkeOut|WWWOFFLE) 1;
  ~*(GrabNet|NetSpider|Vampire|NetZIP|Octopus|Offline|PageGrabber|Foto|pavuk|pcBrowser|RealDownload|ReGet|SiteSnagger|SmartDownload|SuperBot|WebSpider) 1;
  ~*(Teleport|VoidEYE|Collector|WebAuto|WebCopier|WebFetch|WebGo|WebLeacher|WebReaper|WebSauger|eXtractor|Quester|WebZIP|Wget|Widow|Zeus) 1;
  ~*(Twengabot|htmlparser|libwww|Python|perl|urllib|scan|Curl|email|PycURL|Pyth|PyQ|WebCollector|WebCopy|webcraw) 1;
}

Add the URL redirection inside of your server block:

server {
  # server_name, root, etc
  # ...
  location /faceless {
    try_files $uri $uri/ =404;
  }
  location / {
    set $face_matches 0;
    if (-f /faceless$request_filename) {
      set $face_matches 1$limit_bots;
    }
    location ~* \.(jpg|jpeg|png|gif)$ {
      if ($limit_bots = 11) {
        return 301 /faceless$request_uri;
      }
    }
    try_files $uri $uri/ =404;
  }

Apache

• Sample Apache configuration

First, make sure you have mod_rewrite enabled.

sudo a2enmod rewrite

Integrate the following code into the configuration for your domain name:

<Directory "/">
  RewriteEngine on
  RewriteCond %{REQUEST_URI} !^/faceless/
  RewriteCond %{HTTP_USER_AGENT} googlebot|yahoobot|microsoftbot [NC]
  RewriteRule (\.jpg|\.jpeg|\.gif|\.png)$ /faceless%{REQUEST_URI} [L,R=301]
</Directory>

Sources

Our list of crawlers is built from the monperrus/crawler-user-agents.

Prototype Ideas

• explore using adversarial modifications to face image using fawkes style negative-face interpolation
• explore using adversarial noise into facial region
• explore using face resolution degradation techniques to simply lower the amount of unique information in a face while still presenting the same image to bots and users
• develop benchmarking scripts to analyze amount of biometric information on a website

Known Issues

• Using if statements in NGINX is not allowed. Can't use if with sub_filter.
• Referrer and user-agent can and will be easily spoofed
• Google indexes the URL to the non-redacted image and does not seem to follow the 301 redirect. This exposes the unredacted image.
• On Firefox Ubuntu, Firefox sends referrer facetest.vframe.io which reveals the unredacted image, but not on MacOS