Detecting dark web mirror sites on the ClearNET using encrypted traffic classification

Dark Web Mirror sites accessible through ordinary browsers may expose users to phishing attacks, malicious content, and illicit material while falling outside the scope of monitoring approaches focused on Tor traffic. We investigated whether such access could be identified at the session level using packet-direction information alone. We collected 4,600 sessions from 23 Clear Web Uniform Resource Locators (URLs) and 23 Dark Web Mirror site URLs, with 100 repeated 20-second visits to each URL. Traffic retained after Transmission Control Protocol (TCP), Transport Layer Security (TLS), and Quick UDP Internet Connections (QUIC) filtering was represented as fixed-length sequences of client-to-server, server-to-client, and padding symbols, excluding payloads, network addresses, page content, packet sizes, and timing from the model input. We evaluated observation windows, median-derived sequence lengths, removal of the initial 20-packet segment, sequence-length reduction, and ten sequence classifiers. Under session-level stratified 5-fold evaluation with a sequence length of 526 selected from the 3-second observation window, the baseline receiver operating characteristic area under the curve (ROC-AUC) was \(0.9847 \pm 0.0016\) . Under session-level stratified 5-fold evaluation with an input length of 526 including the initial 20 packets, one-dimensional Residual Network (ResNet1D) achieved an accuracy of \(0.9260 \pm 0.0057\) , an F1 score of \(0.9264 \pm 0.0055\) , and a ROC-AUC of \(0.9813 \pm 0.0016\) . These values are the means and sample standard deviations across the five final seeds, and the mean total processing time was 7.6808 ms. In the additional registrable-domain-disjoint Group-aware 5-fold evaluation, ResNet1D achieved a ROC-AUC of \(0.9431 \pm 0.0069\) . Interpretation of this result is restricted to domain-disjoint evaluation within the existing dataset. These results support the feasibility of early session-level detection of access to Dark Web Mirror sites in ordinary browser environments.

Authors

Institutions

Publication Details

Journal
Scientific Reports
Published
2026-11-13
DOI
https://doi.org/10.1038/s41598-026-80334-2
Primary Topic
Internet Traffic Analysis and Secure E-voting
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Detecting dark web mirror sites on the ClearNET using encrypted traffic classification

Joonseo Ha, Sungwook Ryu, Mee Lan Han, Youngjin Park et al.
1 links
Scientific Reports
Internet Traffic Analysis and Secure E-voting
article

Detecting dark web mirror sites on the ClearNET using encrypted traffic classification

Joonseo Ha, Sungwook Ryu, Mee Lan Han, Youngjin Park, Jihun Han, Juhyeok Jeon
article en

Abstract

Dark Web Mirror sites accessible through ordinary browsers may expose users to phishing attacks, malicious content, and illicit material while falling outside the scope of monitoring approaches focused on Tor traffic. We investigated whether such access could be identified at the session level using packet-direction information alone. We collected 4,600 sessions from 23 Clear Web Uniform Resource Locators (URLs) and 23 Dark Web Mirror site URLs, with 100 repeated 20-second visits to each URL. Traffic retained after Transmission Control Protocol (TCP), Transport Layer Security (TLS), and Quick UDP Internet Connections (QUIC) filtering was represented as fixed-length sequences of client-to-server, server-to-client, and padding symbols, excluding payloads, network addresses, page content, packet sizes, and timing from the model input. We evaluated observation windows, median-derived sequence lengths, removal of the initial 20-packet segment, sequence-length reduction, and ten sequence classifiers. Under session-level stratified 5-fold evaluation with a sequence length of 526 selected from the 3-second observation window, the baseline receiver operating characteristic area under the curve (ROC-AUC) was \(0.9847 \pm 0.0016\) . Under session-level stratified 5-fold evaluation with an input length of 526 including the initial 20 packets, one-dimensional Residual Network (ResNet1D) achieved an accuracy of \(0.9260 \pm 0.0057\) , an F1 score of \(0.9264 \pm 0.0055\) , and a ROC-AUC of \(0.9813 \pm 0.0016\) . These values are the means and sample standard deviations across the five final seeds, and the mean total processing time was 7.6808 ms. In the additional registrable-domain-disjoint Group-aware 5-fold evaluation, ResNet1D achieved a ROC-AUC of \(0.9431 \pm 0.0069\) . Interpretation of this result is restricted to domain-disjoint evaluation within the existing dataset. These results support the feasibility of early session-level detection of access to Dark Web Mirror sites in ordinary browser environments.

Scientific Reports
Korea Advanced Institute of Science and Technology (KR), Cyber University (JP), Sejong Cyber University (KR)
Ministry of Trade, Industry and Energy, National Research Foundation of Korea, Ministry of Science and ICT, South Korea, Institute for Information and Communications Technology Promotion
Internet Traffic Analysis and Secure E-voting
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon