Context-Aware Dynamic Pruning (CADP): An Approach for Accelerating Model Training in Website Fingerprinting Attacks
IEEE Access · 2026 · 14 · 166264-166278
DOI: 10.1109/access.2026.3742123 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Mohamad Amar Irsyad Mohd Aminuddin und 3 weiteren Autor:innen. Die Arbeit nennt 39 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0). Gefördert von: Universiti Sains Malaysia.
Abstract (Originalsprache)
Website Fingerprinting (WF) attacks analyse encrypted network traffic metadata to identify webpages browsed over anonymity networks such as Tor. State-of-the-art WF attacks rely on deep learning (DL) models to extract complex traffic patterns. However, fitting these deep neural networks requires iterative training across numerous epochs, creating a computational bottleneck during experimentation and retraining under traffic drift. While existing research focuses on input sequence truncation and model parameter pruning, the number of traces processed in each training epoch remains largely unaddressed. As WF datasets contain repeated visits to the same monitored webpages, processing every well-learned trace in every epoch introduces unnecessary computational repetition, creating a computational bottleneck. To resolve this computational bottleneck, this study proposes Context-Aware Dynamic Pruning (CADP), a training-loop enhancement designed for WF models. CADP dynamically selects the traces processed in each epoch based on their recorded training loss, rescales the gradients of kept traces to preserve unbiased optimisation and enforces a per-webpage coverage floor to prevent minority class degradation. Across seven diverse WF attacks evaluated over six seed values, CADP reduces training time by 10.8% to 30.5% (mean reduction of 21.4% and speedups of 1.12x to 1.44x), with all reductions statistically confirmed under Holm-Bonferroni adjusted tests (p < 0.001). Classification accuracy differences remain within 0.70 percentage points across six evaluated architectures and 1.96 percentage points on HiFi-WF (benchmark mean difference of -0.57 percentage points). By accelerating model fitting without altering neural architectures, CADP provides a drop-in solution that shortens the experimental cycle in WF research.
Förderung
- Universiti Sains Malaysia
Zitieren
APA
Aminuddin, M. A. I. M., Zaaba, Z. F., Talib, S., & Shariff, M. R. M. (2026). Context-Aware Dynamic Pruning (CADP): An Approach for Accelerating Model Training in Website Fingerprinting Attacks. IEEE Access, 14, 166264-166278. https://doi.org/10.1109/access.2026.3742123
Vancouver
Aminuddin MAIM, Zaaba ZF, Talib S, Shariff MRM. Context-Aware Dynamic Pruning (CADP): An Approach for Accelerating Model Training in Website Fingerprinting Attacks. IEEE Access. 2026;14:166264-166278. doi:10.1109/access.2026.3742123
BibTeX
@article{Aminuddin2026,
author = {Aminuddin, Mohamad Amar Irsyad Mohd and Zaaba, Zarul Fitri and Talib, Shuhaili and Shariff, Mohd Ridzuan M.},
title = {Context-Aware Dynamic Pruning (CADP): An Approach for Accelerating Model Training in Website Fingerprinting Attacks},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {166264--166278},
doi = {10.1109/access.2026.3742123}
}
EUVIMED