NVIDIA DGX A100 Security Update

The DGX A100 System Firmware Update container version 20.11.3 for Ubuntu with BMC version 00.13.04 fixes vulnerabilities described in NVIDIA Security Bulletin 5010 such as CVE‑2020‑11487.

More details can be found in recent AISec talks and releases. 

Enjoy

Vulnerabilities of Machine Learning Infrastructure (Slides/Video)

Vulnerabilities of Machine Learning Infrastructure talk as presented at CodeBlue 2020 Japan and Standoff365 by Sergey Gordeychik.

The boom of AI brought to the market a set of impressive solutions both on the hardware and software side. On the other hand, massive implementation of AI in various areas brings about problems, and security is one of the greatest concerns.

Vulnerabilities of Machine Learning Infrastructure

As presented at The Standoff online cyber-range and security conference by Sergey Gordeychik.

In this talk we will present results of hands-on vulnerability research of different components of AI infrastructure including NVIDIA DGX GPU servers, ML frameworks such as Pytorch, Keras and Tensorflow, data processing pipelines and specific applications, including Medical Imaging and face recognition powered CCTV. Updated Internet Census toolkit based on the Grinder framework will be introduced.

NVIDIA DGX machine learning servers vulnerabilities

NVIDIA has published fixes for vulnerabilities in NVIDIA Machine learning servers with CVSS up to 9.8.

NVIDIA DGX-1, DGX-2, and DGX A100 Servers are affected and can be hacked via BMC OOB interfaces. 

codeblue.jp talk

Vulnerabilities of Machine Learning Infrastructure talk to be presented at Code Blue 2020 security conference @Tokyo. 

Vulnerabilities in AI Healthcare pipelines

Must see if you use/develop Artificial Intelligence in Healthcare and care about Cybersecurity and Privacy.

How to make your own Internet Census

Simple writeup on the Internet-scale census with example or Artificial Intelligence and Machine Learning infrastructure assessment by Antony Nikolaev. Sample Lab of Cybersecurity of Machine Learning and Artificial Intelligence at Harbour.Space University.

Just in case if you need spare Tensorboard in Africa or Kubeflow elsewhere.

Hacking Odyssey at HITBLockdown002

How to Hack Medical Imaging Applications via DICOM by Maria Nedyak

DATE: July 25, 2020
TIME: 04:00 PM - 05:00 PM (GMT +8)

Hacking Kubeflow for fun and mining

Microsoft Azure Security Center (ASC) recently published detailed description of Kubeflow backdooring attack.

AI Finger 2020

New release of Internet census of Machine Learning and Artificial Intelligence Frameworks and Applications, April 2020.

Artificial Intelligence Security Census

In this paper, we present the results of Internet-wide security scans of publicly available AIML systems. We show that many different interfaces of AIML systems are not protected and accessible from the Internet, moreover, most of them don't even have basic security mechanisms. Also, we describe found the known vulnerabilities related to outdated software and insecure configurations. 

Malign Machine Learning Models and bad DICOM

Zeronighs 2019 AISec releases: how to insert malware into TensorFlow and PyTorch models and hack NVIDIA Clara ML pipeline with DICOM image.

AISec on ZeroNights 2019

New AISec releases soon at Zeronights 2019.

Roman Palkin : "Malign Machine Learning Models"
Maria Nedyak : "Hacking Medical Imaging with DICOM"


See you there!

Slides and videos

Enjoy 

HITB AISec slides and special release

Slides "AI for Security and Security for AI" talk by Sergey Gordeychik, as presented at HITB CyberWeek 2019, Abu Dhabi.



Machine learning technologies are turning from rocket science into daily engineering life. You no longer have to know the difference between Faster R-CNN and HMM to develop a machine vision system, and even OpenCV has bindings for JavaScript allowing to resolve quite serious tasks all the while remaining in front end. On other hand massive implementation of AI in various areas brings about problems, and security is one of the greatest concerns. In the broader context security is really all about trust.

Do we trust AI? I don’t, personally.

DICOM to passwd. On security of ML pipelines

Machine Learning and Artificial Intelligence Pipelines are very useful tools. They help to concentrate on specific task without digging into implementation details. However, from design and security perspective these things are like Frankenstein.

Here is  an example

Dangers of MLaaS

Dangers of MLaaS as present on Datafest Siberia by  @dnkolegov and Antoniy Nikolaev.

More details in the Measuring Artificial Intelligence and Machine Learning Implementation Security on the Internet writeup.


Download

Enjoy 

AI Finger

Special release for XCON 2019 Beijing conference and “AI for Security or Security for AI” talk by Sergey Gordeychik.

Very first (and very draft) round of Machine Learning and Artificial Intelligence Frameworks and Applications Fingerprinting and Internet Census project.
Based on the Grinder framework.