Research

NLP & Rights Metadata 2024 – present

A deployed NER + LLM “division of labor” pipeline for Korean public-domain rights documents: a multi-provider OCR fallback chain feeds a fine-tuned KLUE-RoBERTa-Large NER model (26 B-I-O labels) and schema-guided LLM extraction running concurrently, and an LLM consolidation arbiter produces a unified 67-field schema with per-field provenance (AGREED / CONFLICT / LLM_ONLY / NER_ONLY / MISSING) and OCR-grounded Korean evidence.

Year 2 adds image + text multimodal analysis — generative VLMs and SigLIP 2 + FAISS similarity — over a 144k-work KOGL corpus. The work is funded by the MCST/KOCCA copyright R&D program, in a consortium with Muhayu and HM Company serving the Korea Copyright Commission (Gongu Madang) and the Korea Culture Information Service Agency (KOGL).

Architecture diagram: OCR feeds concurrent LLM extraction and Korean NER, an LLM arbiter consolidates both into a unified 67-field metadata schema
NER–LLM consolidation architecture for rights-metadata extraction (deployed pipeline)

Voice Phishing Detection 2021 – present

South Korea’s National Police Agency counts 260,694 cumulative vishing incidents from 2016 to 2025, with KRW 6,015.3 billion (≈ USD 4.11 B) in damage. Against that backdrop the lab built KorCCVi, the first labeled Korean voice-phishing transcript dataset, and progressed from ML/DL text classifiers and KoBERT to an attention-based CNN-BiLSTM (F1 0.9966), federated learning, and multilingual back-translation with SMOTE augmentation.

The current line is multimodal (audio + text) and privacy-preserving: lightweight acoustic models (≤223K parameters; 99.59% accuracy with eGeMAPS features) and channel-robust Wav2Vec2.0 with EMA call-level aggregation (F1 ≈ 0.956–0.96 under GSM codec-matched evaluation) — efficient enough for real-time detection; see the papers for benchmark details.

Pipeline diagram: input audio stream through preprocessing, eGeMAPS/MFCC/Wav2Vec2.0 feature extraction, a lightweight MLP classifier, and EMA call-level aggregation to a vishing alert — fully on-device
On-device, privacy-preserving detection pipeline — acoustic features to a call-level alert (IEEE Access line)

Digital Forensics 2010 – 2020

A decade of digital forensics on flash storage: recovery of deleted and wiped files on NAND flash, anti-forensics via block permutation to prevent recovery, wiping-evidence acquisition, selective recovery under the SSD TRIM command, YAFFS2 file-system metadata management, and circumstantial-evidence recovery for forensic timeline analysis.

The same line extends to databases in the field — SQLite deleted-record recovery, journaling-of-journal-based SQLite file recovery, and message recovery from instant messengers. Altogether this area produced more than ten master’s theses between 2010 and 2020.

Diagram of block-mapping and log-page-mapping tables being restored from map blocks and pruned by a consistency check after a crash
Mapping-table restoration and consistency check after power failure (Electronics 2021)

Flash Storage & FTL 2004 – present

The lab co-created the widely cited FAST flash translation layer (“A log buffer-based flash translation layer using fully-associative sector translation”, ACM TECS 2007 — cited 700+ times) and the LSTAFF/LSTAFF* system software for large-block flash. The work spans FTL design and surveys, hybrid address mapping, buffer replacement, flash-aware B-tree indexing (WPCB-Tree), and crash and data recovery on NAND flash and SSDs.

Its roots run deeper still: before flash, the lab worked on database systems and spatial/multimedia indexing — k-nearest-neighbor queries with semantic predicates (Spy-Tec+, RS-tree), distance browsing, and GIS object prefetching (1999–2011). Today the flash line feeds directly into the undergraduate File Processing course, whose assignments build an emulated NAND device driver and an FTL from scratch.

Architecture diagram of the WPCB-tree: a B+-tree module whose random writes pass through a transit buffer and write-pattern converter into sequential writes over the FTL and flash memory
WPCB-tree — a flash-aware B-tree over the FTL (Symmetry 2018)

Datasets & Code

Projects