NFStream is a multiplatform Python framework providing fast, flexible, and expressive data structures designed to make working with online or offline network data easy and intuitive. It aims to be Python's fundamental high-level building block for doing practical, real-world network flow data analysis. Additionally, it has the broader goal of becoming a unifying network data analytics framework for researchers providing data reproducibility across experiments.
Live Notebook |
|
Project Website |
|
Discussion Channel |
|
Latest Release |
|
Supported Versions |
|
Project License |
|
Continuous Integration |
|
Code Quality |
|
Code Coverage |
|
- Main Features
- How to get it?
- How to use it?
- Building from sources
- Contributing
- Ethics
- Credits
- Publications that use NFStream
- License
- Performance: NFStream is designed to be fast: AF_PACKET_V3/FANOUT on Linux, multiprocessing, native CFFI based computation engine, and PyPy full support.
- Encrypted layer-7 visibility: NFStream deep packet inspection is based on nDPI. It allows NFStream to perform reliable encrypted applications identification and metadata fingerprinting (e.g. TLS, SSH, DHCP, HTTP).
- System visibility: NFStream probes the monitored system's kernel to obtain information on open Internet sockets and collects guaranteed ground-truth (process name, PID, etc.) at the application level.
- Statistical features extraction: NFStream provides state of the art of flow-based statistical feature extraction. It includes post-mortem statistical features (e.g., minimum, mean, standard deviation, and maximum of packet size and inter-arrival time) and early flow features (e.g. sequence of first n packets sizes, inter-arrival times, and directions).
- Flexibility: NFStream is easily extensible using NFPlugins. It allows the creation of a new flow feature within a few lines of Python.
- Machine Learning oriented: NFStream aims to make Machine Learning Approaches for network traffic management reproducible and deployable. By using NFStream as a common framework, researchers ensure that models are trained using the same feature computation logic, and thus, a fair comparison is possible. Moreover, trained models can be deployed and evaluated on live networks using NFPlugins.
Binary installers for the latest released version are available on Pypi.
pip install nfstream
Windows Notes: NFStream does not include capture drivers on Windows (license restrictions). It is required to install Npcap drivers before installing NFStream. If Wireshark is already installed on Windows, then Npcap drivers are already installed, and you do not need to perform any action.
Dealing with a big pcap file and want to aggregate into labeled network flows? NFStream make this path easier in a few lines:
from nfstream import NFStreamer
# We display all streamer parameters with their default values.
# See documentation for detailed information about each parameter.
# https://www.nfstream.org/docs/api#nfstreamer
my_streamer = NFStreamer(source="facebook.pcap", # or live network interface
decode_tunnels=True,
bpf_filter=None,
promiscuous_mode=True,
snapshot_length=1536,
idle_timeout=120,
active_timeout=1800,
accounting_mode=0,
udps=None,
n_dissections=20,
statistical_analysis=False,
splt_analysis=0,
n_meters=0,
max_nflows=0,
performance_report=0,
system_visibility_mode=0,
system_visibility_poll_ms=100)
for flow in my_streamer:
print(flow) # print it.
# See documentation for each feature detailed description.
# https://www.nfstream.org/docs/api#nflow
NFlow(id=0,
expiration_id=0,
src_ip='192.168.43.18',
src_mac='30:52:cb:6c:9c:1b',
src_oui='30:52:cb',
src_port=52066,
dst_ip='66.220.156.68',
dst_mac='98:0c:82:d3:3c:7c',
dst_oui='98:0c:82',
dst_port=443,
protocol=6,
ip_version=4,
vlan_id=0,
tunnel_id=0,
bidirectional_first_seen_ms=1472393122365,
bidirectional_last_seen_ms=1472393123665,
bidirectional_duration_ms=1300,
bidirectional_packets=19,
bidirectional_bytes=5745,
src2dst_first_seen_ms=1472393122365,
src2dst_last_seen_ms=1472393123408,
src2dst_duration_ms=1043,
src2dst_packets=9,
src2dst_bytes=1345,
dst2src_first_seen_ms=1472393122668,
dst2src_last_seen_ms=1472393123665,
dst2src_duration_ms=997,
dst2src_packets=10,
dst2src_bytes=4400,
application_name='TLS.Facebook',
application_category_name='SocialNetwork',
application_is_guessed=0,
application_confidence=4,
requested_server_name='facebook.com',
client_fingerprint='bfcc1a3891601edb4f137ab7ab25b840',
server_fingerprint='2d1eb5817ece335c24904f516ad5da12',
user_agent='',
content_type='')
NFStream probes the monitored system's kernel to obtain information on open Internet sockets and collects guaranteed ground-truth (process name, PID, etc.) at the application level.
from nfstream import NFStreamer
my_streamer = NFStreamer(source="Intel(R) Wi-Fi 6 AX200 160MHz", # Live capture mode.
# Disable L7 dissection for readability purpose only.
n_dissections=0,
system_visibility_poll_ms=100,
system_visibility_mode=1)
for flow in my_streamer:
print(flow) # print it.
# See documentation for each feature detailed description.
# https://www.nfstream.org/docs/api#nflow
NFlow(id=0,
expiration_id=0,
src_ip='192.168.43.18',
src_mac='30:52:cb:6c:9c:1b',
src_oui='30:52:cb',
src_port=59339,
dst_ip='184.73.244.37',
dst_mac='98:0c:82:d3:3c:7c',
dst_oui='98:0c:82',
dst_port=443,
protocol=6,
ip_version=4,
vlan_id=0,
tunnel_id=0,
bidirectional_first_seen_ms=1638966705265,
bidirectional_last_seen_ms=1638966706999,
bidirectional_duration_ms=1734,
bidirectional_packets=98,
bidirectional_bytes=424464,
src2dst_first_seen_ms=1638966705265,
src2dst_last_seen_ms=1638966706999,
src2dst_duration_ms=1734,
src2dst_packets=22,
src2dst_bytes=2478,
dst2src_first_seen_ms=1638966705345,
dst2src_last_seen_ms=1638966706999,
dst2src_duration_ms=1654,
dst2src_packets=76,
dst2src_bytes=421986,
# The process that generated this reported flow.
system_process_pid=14596,
system_process_name='FortniteClient-Win64-Shipping.exe')
NFStream performs 48 post-mortem flow statistical features extraction, which includes detailed TCP flags analysis, minimum, mean, maximum, and standard deviation of both packet size and inter-arrival time in each direction.
from nfstream import NFStreamer
my_streamer = NFStreamer(source="facebook.pcap",
# Disable L7 dissection for readability purpose.
n_dissections=0,
statistical_analysis=True)
for flow in my_streamer:
print(flow)
# See documentation for each feature detailed description.
# https://www.nfstream.org/docs/api#nflow
NFlow(id=0,
expiration_id=0,
src_ip='192.168.43.18',
src_mac='30:52:cb:6c:9c:1b',
src_oui='30:52:cb',
src_port=52066,
dst_ip='66.220.156.68',
dst_mac='98:0c:82:d3:3c:7c',
dst_oui='98:0c:82',
dst_port=443,
protocol=6,
ip_version=4,
vlan_id=0,
tunnel_id=0,
bidirectional_first_seen_ms=1472393122365,
bidirectional_last_seen_ms=1472393123665,
bidirectional_duration_ms=1300,
bidirectional_packets=19,
bidirectional_bytes=5745,
src2dst_first_seen_ms=1472393122365,
src2dst_last_seen_ms=1472393123408,
src2dst_duration_ms=1043,
src2dst_packets=9,
src2dst_bytes=1345,
dst2src_first_seen_ms=1472393122668,
dst2src_last_seen_ms=1472393123665,
dst2src_duration_ms=997,
dst2src_packets=10,
dst2src_bytes=4400,
bidirectional_min_ps=66,
bidirectional_mean_ps=302.36842105263156,
bidirectional_stddev_ps=425.53315715259754,
bidirectional_max_ps=1454,
src2dst_min_ps=66,
src2dst_mean_ps=149.44444444444446,
src2dst_stddev_ps=132.20354676701294,
src2dst_max_ps=449,
dst2src_min_ps=66,
dst2src_mean_ps=440.0,
dst2src_stddev_ps=549.7164925870628,
dst2src_max_ps=1454,
bidirectional_min_piat_ms=0,
bidirectional_mean_piat_ms=72.22222222222223,
bidirectional_stddev_piat_ms=137.34994188549086,
bidirectional_max_piat_ms=398,
src2dst_min_piat_ms=0,
src2dst_mean_piat_ms=130.375,
src2dst_stddev_piat_ms=179.72036811192467,
src2dst_max_piat_ms=415,
dst2src_min_piat_ms=0,
dst2src_mean_piat_ms=110.77777777777777,
dst2src_stddev_piat_ms=169.51458475436397,
dst2src_max_piat_ms=409,
bidirectional_syn_packets=2,
bidirectional_cwr_packets=0,
bidirectional_ece_packets=0,
bidirectional_urg_packets=0,
bidirectional_ack_packets=18,
bidirectional_psh_packets=9,
bidirectional_rst_packets=0,
bidirectional_fin_packets=0,
src2dst_syn_packets=1,
src2dst_cwr_packets=0,
src2dst_ece_packets=0,
src2dst_urg_packets=0,
src2dst_ack_packets=8,
src2dst_psh_packets=4,
src2dst_rst_packets=0,
src2dst_fin_packets=0,
dst2src_syn_packets=1,
dst2src_cwr_packets=0,
dst2src_ece_packets=0,
dst2src_urg_packets=0,
dst2src_ack_packets=10,
dst2src_psh_packets=5,
dst2src_rst_packets=0,
dst2src_fin_packets=0)
NFStream performs early (up to 255 packets) flow statistical features extraction (referred to as SPLT analysis in the literature). It is summarized as a sequence of these packets' directions, sizes, and inter-arrival times.
from nfstream import NFStreamer
my_streamer = NFStreamer(source="facebook.pcap",
# We disable l7 dissection for readability purpose.
n_dissections=0,
splt_analysis=10)
for flow in my_streamer:
print(flow)
# See documentation for each feature detailed description.
# https://www.nfstream.org/docs/api#nflow
NFlow(id=0,
expiration_id=0,
src_ip='192.168.43.18',
src_mac='30:52:cb:6c:9c:1b',
src_oui='30:52:cb',
src_port=52066,
dst_ip='66.220.156.68',
dst_mac='98:0c:82:d3:3c:7c',
dst_oui='98:0c:82',
dst_port=443,
protocol=6,
ip_version=4,
vlan_id=0,
tunnel_id=0,
bidirectional_first_seen_ms=1472393122365,
bidirectional_last_seen_ms=1472393123665,
bidirectional_duration_ms=1300,
bidirectional_packets=19,
bidirectional_bytes=5745,
src2dst_first_seen_ms=1472393122365,
src2dst_last_seen_ms=1472393123408,
src2dst_duration_ms=1043,
src2dst_packets=9,
src2dst_bytes=1345,
dst2src_first_seen_ms=1472393122668,
dst2src_last_seen_ms=1472393123665,
dst2src_duration_ms=997,
dst2src_packets=10,
dst2src_bytes=4400,
# The sequence of 10 first packet direction, size and inter arrival time.
splt_direction=[0, 1, 0, 0, 1, 1, 0, 1, 0, 1],
splt_ps=[74, 74, 66, 262, 66, 1454, 66, 1454, 66, 463],
splt_piat_ms=[0, 303, 0, 0, 313, 0, 0, 0, 0, 1])
NFStream natively supports Pandas as an export interface.
# See documentation for more details.
# https://www.nfstream.org/docs/api#pandas-dataframe-conversion
from nfstream import NFStreamer
my_dataframe = NFStreamer(source='teams.pcap').to_pandas()[["src_ip",
"src_port",
"dst_ip",
"dst_port",
"protocol",
"bidirectional_packets",
"bidirectional_bytes",
"application_name"]]
my_dataframe.head(5)
NFStream natively supports CSV file format as an export interface.
# See documentation for more details.
# https://www.nfstream.org/docs/api#csv-file-conversion
flows_count = NFStreamer(source='facebook.pcap').to_csv(path=None,
columns_to_anonymize=(),
flows_per_file=0,
rotate_files=0)
Didn't find a specific flow feature? add a plugin to NFStream in a few lines:
from nfstream import NFPlugin
class MyCustomPktSizeFeature(NFPlugin):
def on_init(self, packet, flow):
# flow creation with the first packet
if packet.raw_size == self.custom_size:
flow.udps.packet_with_custom_size = 1
else:
flow.udps.packet_with_custom_size = 0
def on_update(self, packet, flow):
# flow update with each packet belonging to the flow
if packet.raw_size == self.custom_size:
flow.udps.packet_with_custom_size += 1
extended_streamer = NFStreamer(source='facebook.pcap',
udps=MyCustomPktSizeFeature(custom_size=555))
for flow in extended_streamer:
# see your dynamically created metric in generated flows
print(flow.udps.packet_with_custom_size)
The following simplistic example demonstrates how to train and deploy a machine-learning approach for traffic flow categorization. We want to run a classification of Social Network category flows based on bidirectional_packets and bidirectional_bytes as input features. For the sake of brevity, we decide to predict only at the flow expiration stage.
from nfstream import NFPlugin, NFStreamer
import numpy
from sklearn.ensemble import RandomForestClassifier
df = NFStreamer(source="training_traffic.pcap").to_pandas()
X = df[["bidirectional_packets", "bidirectional_bytes"]]
y = df["application_category_name"].apply(lambda x: 1 if 'SocialNetwork' in x else 0)
model = RandomForestClassifier()
model.fit(X, y)
class ModelPrediction(NFPlugin):
def on_init(self, packet, flow):
flow.udps.model_prediction = 0
def on_expire(self, flow):
# You can do the same in on_update entrypoint and force expiration with custom id.
to_predict = numpy.array([flow.bidirectional_packets,
flow.bidirectional_bytes]).reshape((1,-1))
flow.udps.model_prediction = self.my_model.predict(to_predict)
ml_streamer = NFStreamer(source="eth0", udps=ModelPrediction(my_model=model))
for flow in ml_streamer:
print(flow.udps.model_prediction)
More NFPlugin examples and details are provided in the official documentation. You can also test NFStream without installation using our live demo notebook.
To build NFStream from sources, please read the installation guide provided in the official documentation.
Please read Contributing for details on our code of conduct and the process for submitting pull requests to us.
NFStream is intended for network data research and forensics. Researchers and network data scientists can use this framework to build reliable datasets and train and evaluate network-applied machine learning models. As with any packet monitoring tool, NFStream could be misused. Do not run it on any network that you do not own or administrate.
NFStream paper is published in Computer Networks (COMNET). If you use NFStream in a scientific publication, we would appreciate citations to the following article:
@article{AOUINI2022108719,
title = {NFStream: A flexible network data analysis framework},
author = {Aouini, Zied and Pekar, Adrian},
doi = {10.1016/j.comnet.2021.108719},
issn = {1389-1286},
journal = {Computer Networks},
pages = {108719},
year = {2022},
publisher = {Elsevier},
volume = {204},
url = {https://www.sciencedirect.com/science/article/pii/S1389128621005739}
}
The following people contributed to NFStream:
- Zied Aouini: Creator and core developer.
- Adrian Pekar: Datasets generation and storage.
- Romain Picard: MDNS and DHCP plugins implementation.
- Radion Bikmukhamedov: Initial work on SPLT analysis NFPlugin.
The following organizations supported NFStream:
- SoftAtHome: Supporter of NFStream development.
- Technical University of Košice: Hardware and infrastructure for datasets generation and storage.
- ntop: Technical support of nDPI integration.
- The Nmap Project: Technical support of Npcap integration (NPCAP OEM installer on Windows CI).
- Google OSS Fuzz: Continious fuzzing testing support of NFStream project.
- A Hierarchical Architecture and Probabilistic Strategy for Collaborative Intrusion Detectionn
- Open-Source Framework for Encrypted Internet and Malicious Traffic Classification
- ConFlow: Contrast Network Flow Improving Class-Imbalanced Learning in Network Intrusion Detection
- Continual Learning for Anomaly based Network Intrusion Detection
- A self-secure system based on software-defined network
- Robust Variational Autoencoders and Normalizing Flows for Unsupervised Network Anomaly Detection
- RADON: Robust Autoencoder for Unsupervised Anomaly Detection
- A Generic Machine Learning Approach for IoT Device Identification
- Ranking Network Devices for Alarm Prioritisation: Intrusion Detection Case Study
- Network Flows-Based Malware Detection Using A Combined Approach of Crawling And Deep Learning
- Network Intrusion Detection Based on Distributed Trustworthy Artificial Intelligence
- Generative Transformer Framework For Network Traffic Generation And Classification
- Multi-Class Network Traffic Generators and Classifiers Based on Neural Networks
- Using Embedded Feature Selection and CNN for Classification on CCD-INID-V1 A New IoT Dataset
- An Approach Based on Knowledge-Defined Networking for Identifying Video Streaming Flows in 5G Networks
- Knowledge Discovery: Can It Shed New Light on Threshold Definition for Heavy‑Hitter Detection?
- Collecting and analyzing Tor exit node traffic
- Analysis and Collection Data from IP Network
This project is licensed under the LGPLv3 License - see the License file for details