GoDaddy Multi-Year Breach || Backdoors in ML
CyberInsights #83 - GoDaddy? || Machine learning models can have backdoors too
The $ 4Bn domain registrar breached over the years
Small businesses should stand up and take notice
GoDaddy, one of the largest domain registrars has reported a multi year breach where hackers were able to access the cPanel of websites and redirect some of them to malicious links. The cPanel is the management url for GoDaddy sites where you can login as admin and make changes. Read more about cPanel here.
There were 2 other incidents reported over the years.
Take Action:
Review all your GoDaddy domains and check for any unknown redirects. Check if your DNS records are being used to point to some other domains or IPs. Do a quick health check of all domains. I use MXToolbox.
Backdoors in ML models?
The ‘learning’ part of Machine Learning is susceptible to attacks.
While Wall.E like AI might be a bit in the future, machines today learn quickly.
The way machines learn is by going through known and classified data and adding it to their repository. Show a large number of dogs already classified as dogs, then the model learns to classify any image it receives as dog or not dog. What if I want to hack an ML model to falsely recognise a cat as a dog? I have to re-train the model by showing it hundreds of images of cats falsely classified as dogs.
Researchers have found a way to sneak in a backdoor into the learning phase of ML models where just small changes (called perturbations) can cause the model to learn something completely different. It’s a very interesting development. The earliest known backdoors in ML.
I have spoken of ML hacks earlier as well:
Take Action:
While there is no direct action as of now, we need to learn more about the ML models in use in our organisations. Again, our threat models should consider ML attacks.