Add the concept of "global context". Right now every instance of `struct ndpi_detection_module_struct` (we will call it "local context" in this description) is completely independent from each other. This provide optimal performances in multithreaded environment, where we pin each local context to a thread, and each thread to a specific CPU core: we don't have any data shared across the cores. Each local context has, internally, also some information correlating **different** flows; something like: ``` if flow1 (PeerA <-> Peer B) is PROTOCOL_X; then flow2 (PeerC <-> PeerD) will be PROTOCOL_Y ``` To get optimal classification results, both flow1 and flow2 must be processed by the same local context. This is not an issue at all in the far most common scenario where there is only one local context, but it might be impractical in some more complex scenarios. Create the concept of "global context": multiple local contexts can use the same global context and share some data (structures) using it. This way the data correlating multiple flows can be read/write from different local contexts. This is an optional feature, disabled by default. Obviously data structures shared in a global context must be thread safe. This PR updates the code of the LRU implementation to be, optionally, thread safe. Right now, only the LRU caches can be shared; the other main structures (trees and automas) are basically read-only: there is little sense in sharing them. Furthermore, these structures don't have any information correlating multiple flows. Every LRU cache can be shared, independently from the others, via `ndpi_set_config(ndpi_struct, NULL, "lru.$CACHE_NAME.scope", "1")`. It's up to the user to find the right trade-off between performances (i.e. without shared data) and classification results (i.e. with some shared data among the local contexts), depending on the specific traffic patterns and on the algorithms used to balance the flows across the threads/cores/local contexts. Add some basic examples of library initialization in `doc/library_initialization.md`. This code needs libpthread as external dependency. It shouldn't be a big issue; however a configure flag has been added to disable global context support. A new CI job has been added to test it. TODO: we should need to find a proper way to add some tests on multithreaded enviroment... not an easy task... *** API changes *** If you are not interested in this feature, simply add a NULL parameter to any `ndpi_init_detection_module()` calls.
DGA detection testing workflow
Overview
nDPI provides a set of threat detection features available through NDPI_RISK detection.
As part of these features, we provide DGA detection.
Domain generation algorithms (DGA) are algorithms seen in various families of malware that are used to periodically generate a large number of domain names that can be used as rendezvous points with their command and control servers.
DGA detection heuristic is implemented here.
DGA performances test and tracking allows us to detect automatically if a modification is harmful.
The modification can be a simple threshold change or a future lightweight ML approach.
Used data
Original used dataset is a collection of legit and DGA domains (balanced) that can be obtained as follow:
wget https://raw.githubusercontent.com/chrmor/DGA_domains_dataset/master/dga_domains_full.csv
We split the dataset into DGA and NON-DGA and we keep 10% of each as test set and 90% as training set.
python3 -m pip install pandas
python3 -m pip install sklearn
Instruction using python3
from sklearn.model_selection import train_test_split
import pandas as pd
df = pd.read_csv("dga_domains_full.csv", header=None, names=["type", "family", "domain"])
df_dga = df[df.type=="dga"]
df_non_dga = df[df.type=="legit"]
train_non_dga, test_non_dga = train_test_split(df_non_dga, test_size=0.1, shuffle=True, random_state=27)
train_dga, test_dga = train_test_split(df_dga, test_size=0.1, shuffle=True, random_state=27)
test_dga["domain"].to_csv("test_dga.csv", header=False, index=False)
test_non_dga["domain"].to_csv("test_non_dga.csv", header=False, index=False)
train_dga["domain"].to_csv("train_dga.csv", header=False, index=False)
test_non_dga["domain"].to_csv("test_non_dga.csv", header=False, index=False)
Detection approach must be built on top of training set only, test set must be kept as unseen cases for testing
dga_evaluate
After nDPI compilation, you can use dga_evaluate helper to check number of detections out of an input file.
dga_evaluate <file name>
You can evaluate your modifications performances before submitting it as follows:
./do-dga.sh
If your modifications decreases baseline performances, test will fails. If not (well done), test passed and you must update the baseline metrics with your obtained ones.