Awesome
Dataset, and code for Accurate Mobile-App Fingerprinting Using Flow-level Relationship with Graph Neural Networks.
This work has been accepted by the journey of Computer Networks.
Background of this dataset
We collected another private encrypted mobile application traffic dataset across weeks, to evaluate the generalization of our method in dealing with ambiguous traffic and the performance against traffic concepts drift. Here go the details of the dataset setup:
-
Equipment setup As Figure-3 indicates, smartphones with apps communicate with the Internet via a WiFi access point (AP) and the AP forwards the packets into two gateways that come from different ISPs. To generate traffic from apps, we used scripts that communicated with the target mobile via USB using Android Debug Bridge (ADB). These scripts were sent by the controller computer, and mainly contained UI commands that simulated user actions within apps and system commands that configured the devices.
-
Applications selection We selected 53 apps from the apps list used by AppScanner after filtering out several plain-text applications which consist of a relatively low fraction of encrypted traffic. These apps come from different regions such as shopping, magazines, social and so on. We always installed the latest versions onto the selected devices and signed up for each app.
-
Network trace collection We cyclically performed UI fuzzing operations on each app which is activated for about 30 seconds every time via monkeyrunner as Appscanner and passively collected the network traffic between smartphones and the AP. To collect pure network traces from specific apps, we configured the android devices with the iptables rules and listened on the NFLOG. We also filtered out retransmit, out of order and zero-payload packets.
We collected our private dataset from 23rd June and obtained a dataset named D1. Then we collected dataset D2 one month later after the D1 was captured, and 22 apps have been updated.
Directory Structure
We have extracted the side-channel informantion such as packet size, packet arrival time from the raw pcap files and shared them in the format of json file.
All the json files are layouted as dataset/{datasetName}/{appName}/{appVersion}/{timestamp}_clear.pcap.json.
Let us take an example, for a pcap sample for bbc.mobile.weather of version 4.0.6, the parsed json file is:
dataset/D1/bbc.mobile.weather/4.0.6/1592638931_clear.pcap.json.
Data structure for each Json file
For each parsed json file, it contains all the neccessary side-channel information of all the network flows in its corresponding pcap file.
Here is an example: there are six network flows in the corresponding pcap, so there are six items within the list. For each item, we present the packet length sequence, arrive_time_delta, start_timestamp of flow, end_timestamp of flow and src_port.
[
{
"packet_length": [
282,
1448,
1448,
1448,
1448,
1448,
1448,
1448,
1448,
1448,
1448,
1448,
1448,
1448,
1448,
1448,
1448,
1448,
1448,
1448,
475
],
"payload": [],
"arrive_time_delta": [
0,
0.0000030994415283203125,
0.0002589225769042969,
0.00014519691467285156,
0.0002589225769042969,
0.00004792213439941406,
0.0001900196075439453,
0.0000030994415283203125,
0.0017788410186767578,
0.000010013580322265625,
0.00045800209045410156,
0.0000059604644775390625,
0.000014066696166992188,
0.00005698204040527344,
0.000013113021850585938,
0.000033855438232421875,
0.0047261714935302734,
0.000003814697265625,
0.00006699562072753906,
0.0000030994415283203125,
0.03981494903564453
],
"start_timestamp": 1592747812.358703,
"end_timestamp": 1592747812.406596,
"src_port": 38075
},
{
"packet_length": [
178,
-1448,
-600,
-792,
-1428,
-1021,
126,
53,
44,
42,
-327,
870,
-38,
38,
-728,
-77,
31
],
"payload": [],
"arrive_time_delta": [
0,
0.08965301513671875,
0.00012612342834472656,
0.0000059604644775390625,
0.0989830493927002,
0.0001308917999267578,
0.0000059604644775390625,
0.2170701026916504,
0.000015020370483398438,
0.0000050067901611328125,
0.0001468658447265625,
0.0000059604644775390625,
0.0000050067901611328125,
0.0000030994415283203125,
0.2625288963317871,
13.7536141872406,
0.000007867813110351562
],
"start_timestamp": 1592747813.855156,
"end_timestamp": 1592747828.277463,
"src_port": 43551
},
{
"packet_length": [
174,
-1424,
-1440,
-1448,
-600,
-347,
126,
53,
86,
1448,
524,
-327,
-38,
38,
1448,
1415,
1448,
908,
1448,
328
],
"payload": [],
"arrive_time_delta": [
0,
0.0028209686279296875,
0.00014710426330566406,
0.022533893585205078,
0.0001399517059326172,
0.0000059604644775390625,
0.06827020645141602,
0.0000059604644775390625,
0.00012803077697753906,
0.08455181121826172,
0.000014066696166992188,
0.0001780986785888672,
0.0000059604644775390625,
0.05992698669433594,
0.05175304412841797,
0.00013899803161621094,
1.9377648830413818,
0.00013017654418945312,
2.415329933166504,
0.00000286102294921875
],
"start_timestamp": 1592747814.759279,
"end_timestamp": 1592747819.403128,
"src_port": 37249
},
{
"packet_length": [
174,
-1448,
-600,
-816,
-1440,
-955,
126,
53,
86,
-327
],
"payload": [],
"arrive_time_delta": [
0,
0.14483189582824707,
0.0002219676971435547,
0.008244037628173828,
0.000011920928955078125,
0.036933183670043945,
0.00014090538024902344,
0.0000050067901611328125,
0.0000059604644775390625,
0.411592960357666
],
"start_timestamp": 1592747814.853192,
"end_timestamp": 1592747815.45518,
"src_port": 37250
},
{
"packet_length": [
218,
-1448,
-600,
-750,
93,
-361,
93,
38,
436,
-38,
-1070,
-1440,
-1378,
-1440,
-756,
-268,
46
],
"payload": [],
"arrive_time_delta": [
0,
0.010441064834594727,
0.00015211105346679688,
0.0000059604644775390625,
0.17130398750305176,
0.000011920928955078125,
0.0000040531158447265625,
0.0000019073486328125,
0.0001289844512939453,
0.000008106231689453125,
0.00156402587890625,
0.000011920928955078125,
0.00012612342834472656,
0.0000050067901611328125,
0.6461899280548096,
0.00015592575073242188,
0.0000059604644775390625
],
"start_timestamp": 1592747823.029281,
"end_timestamp": 1592747823.859398,
"src_port": 44086
},
{
"packet_length": [
282,
1448,
1448,
1448,
1448,
1448,
1448,
1448,
1448,
1448,
1448,
1448,
1448,
1448,
1448,
1448,
1448,
1448,
1448,
1448,
1448,
1448,
764
],
"payload": [],
"arrive_time_delta": [
0,
0.00006008148193359375,
0.000102996826171875,
0.000010967254638671875,
0.00003695487976074219,
0.00008392333984375,
0.0000030994415283203125,
0.00009202957153320312,
0.00000286102294921875,
0.010506153106689453,
0.000012874603271484375,
0.000016927719116210938,
0.0002791881561279297,
0.000011920928955078125,
0.0000400543212890625,
0.000009059906005859375,
0.0000059604644775390625,
0.00003790855407714844,
0.0000050067901611328125,
0.00003695487976074219,
0.0000050067901611328125,
0.011543989181518555,
0.0001361370086669922
],
"start_timestamp": 1592747829.196964,
"end_timestamp": 1592747829.220004,
"src_port": 38255
}
]
Currently, we open all the necessary side-channel information of our dataset for convenience, and you can build up Flow relationship graph (FRG) based on these information.
If you want to get the raw pcap files of our dataset, please contact us with the following E-mail: jiangminghao@iie.ac.cn, caiwei@iie.ac.cn, gougaopeng@iie.ac.cn.
If you use our data, or our open source model comparison library, please add a citation about our paper in your paper.
Here goes the bibtex of our paper:
@article{JIANG2022109309,
title = {Accurate mobile-app fingerprinting using flow-level relationship with graph neural networks},
journal = {Computer Networks},
pages = {109309},
year = {2022},
issn = {1389-1286},
doi = {https://doi.org/10.1016/j.comnet.2022.109309},
url = {https://www.sciencedirect.com/science/article/pii/S1389128622003577},
author = {Minghao Jiang and Zhen Li and Peipei Fu and Wei Cai and Mingxin Cui and Gang Xiong and Gaopeng Gou},
keywords = {Mobile encrypted traffic classification, Graph neural network},
abstract = {Identifying mobile applications (apps) from encrypted network traffic (also known as app fingerprinting) plays an important role in areas like network management, advertising analysis, and quality of service. Existing methods mainly extract traffic features from packet-level information (e.g. packet size sequence) and build up classifiers to obtain good performance. However, the packet-level information suffers from small discrimination for the common traffic across apps (e.g. advertising traffic) and rapidly changing for the traffic before and after apps’ updating. As a result, their performance declines in these two real scenes. In this paper, we propose FG-Net, a novel app fingerprinting based on graph neural network (GNN). FG-Net leverages a novel kind of information: flow-level relationship, which is distinctive between different apps and stable across apps’ versions. We design an information-rich graph structure, named FRG, to embed both raw packet-level information and flow-level relationship of traffic concisely. With FRG, we transfer the problem of mobile encrypted traffic fingerprinting into a task of graph representation learning, and we designed a powerful GNN-based traffic fingerprint learner. We conduct comprehensive experiments on both public and private datasets. The results show the FG-Net outperforms the SOTAs in classifying traffic with about 18% common traffic. Without retraining, FG-Net obtains the most robustness against the updated traffic and increases the accuracy by 5.5% compared with the SOTAs.}
}