Skip to main content

Counter-offensive solutions to machine learning

In January 2020 I learned about the results of a contest at DEFCON 2019 where hackers were challenged to come up with counter-offensive solutions to machine learning. One prize winner proved that machine learning could be 100% evaded by mimicking the behaviors of software that produce whitelisted events, and further proved that any whitelisting in machine learning is a vulnerability.

Simply removing the capability to perform whitelisting on an endpoint usually doesn't change the base code of software that uses machine learning since whitelisting (or ignoring patterns found in machine data from an endpoint) is done prior to ingesting machine data used for training. Alternatively I would argue that any kind of function that ignores input or filters input (like regex) from a user is whitelisting.

Do you have a suggestion about how to improve this blog? Let's talk about it. Contact me at DavidBrennerJr@Gmail.com or 720-584-5229.

Comments

Popular posts from this blog

The meaning of time in reinforcement learning

Reinforcement learning (RL) is one of three basic machine learning paradigms, alongside supervised learning and unsupervised learning. Reinforcement learning is concerned with how software agents ought to take actions in an environment in order to maximize the notion of cumulative reward through the process of trial and error. In reinforcement learning an agent starts at an empty state then analyzes the available datasets according to a policy of positive states and negative states. Rather than being explicitly taught as in supervised learning the correct set of actions for performing a task, reinforcement learning uses rewards as signals for positive states and punishments as signals for negative states. The agent obtains the best path to a desirable reward as a cumulation of positive states and negative states. As compared to unsupervised learning, reinforcement learning is different in terms of goals. While the goal in unsupervised learning is to find similarities and differences...

Threat detection of zero-day attacks in Linux with Python

You can investigate suspicious activity that could be a zero-day attack by collecting relevant machine data from your endpoint. You can use the machine data to create your own analysis. Before you start your investigation you will need to determine normal activity on your endpoint. Normal activity is the scope of functionality of the software on your endpoint during periods of low activity and high activity. You will need some kind of software that periodically collects specific machine data from your endpoint like my software developed in Python that's available for free download at https://github.com/davidbrennerjr/server-stats-collector Ingest one or more of the following machine data: General system-wide error messages from /var/log/syslog Auditing logs of application rulesets Auditing logs of security contexts Auditing logs of login attempts from /var/log/auth.log Auditing logs of user management or group manageme...

Web app comparison of async and real-time

Advantages of asynchronous web apps Generic request/response structure Stateless session control Message queue management Token access based on serverside date-time format PostgreSQL paging using token-centric tables and functions Shared pools of resources per customer One-to-many security policies Single domain name with TLS cert over HTTPS Shared bandwidth for uploads/downloads No endpoint/device registration No direct access to server resources Web app aggregation as control panel A/B Testing Advantages of (near) real-time web apps Stateful session control On-demand communication protocols per customer Custom request/response structure per customer Custom date-time formats per customer Endpoint/device registration PostgreSQL paging using static tables and aggregate functions Immediate execution of requests Dedicated pools of resources per customer Dedicated TLS cert over HTTPS per customer Dedicated IP addres...