## PoC [the code](https://github.com/ClonedOne/poisoning_network_flow_classifiers) tested against data collected by [zeek](https://zeek.org/) but likely transferable. by [cloned_tweets et.al](https://twitter.com/cloned_tweets) ## Details Network monitoring involves the collection and analysis of large quantities of security logs ML models for rapid decision making: - To detect potential security threats E.g., botnet detection - To monitor which applications are communicating on the network Often features are composed of aggregated statistics on traffic flows. An adversary that can control a host (or a few hosts) on the network and produce adversarial connection patterns, if included in the training set of a classifier, can poison the training process to ensure a certain behavior is learned by the model, and then a certain trigger pattern can be associated with a certain class of traffic, such as benign. [paper](https://arxiv.org/abs/2306.01655) [slides](https://www.acsac.org/2023/files/web/slides/severi-156-networkpoisoning.pdf)