AI Sec Watch

A bug was found in the Linux kernel before version 5.7.3 in the get_user_pages function (a mechanism that allows programs to access memory pages), where it incorrectly grants write access when it should only allow read access for copy-on-write pages (memory regions shared between processes that are copied when modified). This happens because the function doesn't properly respect read-only restrictions, creating a security vulnerability.

This is an index page summarizing a series of blog posts about machine learning security from a red teaming perspective (testing a system by simulating attacker behavior). The posts cover ML basics, threat modeling, practical attacks like adversarial examples (inputs designed to fool AI models), model theft, backdoors (hidden malicious code inserted into models), and how traditional security attacks (like weak access control) also threaten AI systems.

This post describes how Generative Adversarial Networks (GANs, a type of AI system where two neural networks compete to create realistic fake images) can be used to generate fake husky photos that trick an image recognition system called Husky AI into misclassifying them as real huskies. The author explains they investigated this attack method and references a GAN course to learn more about the technique.

AI and machine learning systems have caused serious problems in real-world situations, including Amazon's recruiting tool that discriminated against women, Microsoft's chatbot that became racist and sexist, IBM's cancer treatment recommendation system that doctors criticized, and Facebook's AI that made incorrect translations leading to someone's arrest. These examples show that AI systems can develop and spread biased predictions and failures with harmful consequences. The article highlights the importance of addressing bias when building and deploying AI systems responsibly.

A vulnerability in Libsvm v324 (a machine learning library used by scikit-learn 0.23.2) allows attackers to crash a program by sending a specially crafted machine learning model with an extremely large value in the _n_support array, causing a segmentation fault (a type of crash where the program tries to access memory it shouldn't). The scikit-learn developers noted this only happens if an application violates the library's API by modifying private attributes.

Repudiation is a security threat where someone denies performing an action, such as replacing an AI model file with a malicious version. The source explains how to use auditd (a Linux auditing tool) and centralized monitoring systems like Splunk or Elastic Stack to create audit logs that track who accessed or modified files and when, helping prove or investigate whether specific accounts made changes.

This is a YouTube talk about building and breaking machine learning systems, presented at a security conference (GrayHat Red Team Village). The speaker is exploring whether to develop this content into a hands-on workshop where participants could practice these concepts.

This post introduces image scaling attacks, a type of adversarial attack (manipulating inputs to fool AI systems) that targets machine learning models through image preprocessing. The author discovered this attack concept while preparing demos and references academic research on understanding and preventing these attacks.

This post demonstrates how to use the Adversarial Robustness Toolbox (ART, an open-source library created by IBM for testing machine learning security) to generate adversarial examples, which are modified images designed to trick AI models into making wrong predictions. The author uses the FGSM attack (Fast Gradient Sign Method, a technique that slightly alters pixel values to confuse classifiers) to successfully manipulate an image of a plush bunny so a husky-recognition AI misclassifies it as a husky with 66% confidence.