Publications & Projects

A record of my research — both completed work and projects currently in progress. Underlined name indicates my authorship.

Publications

Scalable Tor Traffic Correlation Attack via Approximate Nearest Neighbor Search
Saeyeon Hong, Hyewon Kim, Saidur Rahman Mohammad, Se Eun Oh
CISC-S (in Korean), 2026. [pdf]
State-of-the-art Tor traffic correlation methods like DeepCoFFEA (DCF) rely on direct pairwise comparison, incurring O(n²) cost that limits their use at realistic scale. Analyzing the GTT dataset, we estimate an end-to-end adversary must process up to ~196,000 flows per day, and apply CAGRA, a GPU-accelerated approximate nearest neighbor search algorithm, to Tor traffic correlation. Our method preserves or improves detection performance over DCF while substantially cutting runtime and cost, making the attack feasible at real-world scale.

Presentations

[LLM Guided Adversarial Feature-to-Executable Malware Generation]
Saeyeon Hong, Hyejin Woo, Md Mahmuduzzaman Kamol, Se Eun Oh, Md Saidur Rahman
ACSAC, 2025. [poster] [paper]

Projects

Scalable Tor Traffic Correlation Attack via Approximate Nearest Neighbor Search

• In progress ·

With Se-Eun Oh and collaborators.

Reframed pairwise matching in Tor end-to-end correlation attacks as an approximate nearest neighbor search task
Reduced complexity from O(N²) to O(N log N) using GPU-accelerated CAGRA and HNSW indexing
Analyzed the Genuine Tor Traces dataset to quantify the realistic traffic scale an adversary must handle
Proposed a traffic-scale-focused threat model for Tor traffic correlation attacks

Duration Loss as a privacy proxy in Text-to-Speech Model

With Bona Kim and Se-Eun Oh. Second-author project.

Investigated whether duration prediction loss leaks training membership in TTS models
Conducted membership inference attacks using speaker embeddings and alignment-based duration signals
Surveyed the evolution of TTS architectures across autoregressive, non-autoregressive, and duration-based models

KV Cache in AutoPyTorch

With Difan Deng, AutoML Lab, Leibniz University of Hannover (in Germany)

Implemented key–value caching for self-attention blocks in AutoPyTorch decoders (Transformer, TCN, and LSTM)
Performed computational complexity analysis of the caching mechanism, reducing time complexity from O(N²) to O(N) for Transformers and from O(2^L) to O(L) for TCN-based decoders