Kun Woo Park

Kun Woo Park

박건우

I build AI systems that hold up under attack1, and I use AI to take repetitive work out of security2.

  1. SARO, trusted offloading for edge AI, and AgentTrace, which stops risky AI-agent actions before they run.
  2. A compliance LLM that quotes its evidence, and a deepfake detector that writes the takedown request for the victim.
Now
M.S. researcher at the Information Security Lab, Sangmyung University, advised by Prof. Daehee Seo.
Building
TakedownPilot, an AI analyst that verifies phishing and scam reports and drafts the takedown requests.
Elsewhere
GitHub, LinkedIn

Selected work

Aug – Sep 2026

DACON competition

5th on the private leaderboard, Macro F1 0.77455

Public procurement compliance LLM

Reads a public bid notice and decides, for each of 24 legal-violation items, whether the notice breaks the rule, quoting the exact sentence that proves it.

Fixed model (gemma-4-26B-A4B-it), no fine-tuning, no internet, and two hours of inference for the whole test set.

  • vLLM pipeline on one L40S that groups related items into a few prompts per notice and checks that every quote is a verbatim substring of the notice.
  • Per-item decisions that combine the model's judgment with rules from the statute package. On dev, removing a false positive was worth about 4.5 times more F1 than catching a miss, so each item's gate is tuned for precision.
  • Self-labeled a sample of the 20,000 unlabeled notices with frontier LLMs.
  • Paired bootstrap tests on dev to separate real gains from run-to-run noise. The result advanced the team to the final evaluation round.

Python, vLLM, gemma-4-26B-A4B-it

2025 – 2026

CISC-S'25, KIISC

Excellent Paper Award

Registered software, Korea Copyright Commission

Deepfake detection and takedown requests

A victim registers their face once. The system finds manipulated images of them online, confirms they are deepfakes, and sends the takedown request on their behalf.

  • Face matching with MTCNN and InceptionResnetV1, then a ResNet50 and Xception ensemble to judge manipulation. In the paper the ensemble reached 0.830 top-1 accuracy, against 0.770 for ResNet and 0.790 for Xception alone.
  • Takedown emails go out automatically in the site's language. If no reply comes in time, the case is forwarded to the national cybercrime reporting system (ECRM).
  • Victim identity is stored only as SHA-256 hashes, and collected evidence is encrypted with AES-256-GCM.
  • In 2026 I led the team that turned it into a full dashboard for the ICT Challenge, with FastAPI and Celery workers.

Python, PyTorch, MTCNN, ResNet50, Xception, FastAPI, Celery

To

Subject Request for Deepfake Image Removal

Dear Sir/Madam,

I am the administrator of the Deepfake Detection and Removal Request System for victims of Deepfake abuse. We have identified a Deepfake image on your website. Our system kindly requests that you promptly remove the image in question.

If we do not receive a response within one week, we plan to report your website to the cybercrime reporting system.

The English request template from the paper. The same request is generated in other languages, such as Vietnamese.

2026

Personal security tool for Windows

0.08 ms average gate latency

AgentTrace

Stops an AI agent's risky action before it runs, and traces a self-replicating prompt injection back to the first infected file and everything it reached.

  • Rules, fingerprints and a provenance graph make every decision, and the LLM only explains it. Everything works with no network and no LLM.
  • A pre-execution gate answers allow, ask or block for each action, adding about 0.08 ms on average in measurement.
  • Tamper-evident log in SQLite with a hash chain, Sysmon event collection, and a read-only MCP server with four tools. Hooks into Claude Code.
  • Ships as a Windows app with a demo mode, so people who do not code can use it.

Python, SQLite, Sysmon, MCP, pywebview

AgentTrace dashboard in Korean reporting two threats: a self-replicating worm that started in an email and spread to three places, with one action blocked.
Dashboard after a demo attack: a worm that started in an email reached three places, and one action was blocked.

2026

8th KB AI Challenge

Practitioners' Pick #4

Financial consumer protection agent

Helps a bank customer with a complaint: it finds the law and past dispute rulings that apply, shows how similar cases ended, and drafts the application.

  • Runs on the bank's own machines with a small EXAONE model (1.2B, 4-bit) through Ollama, so complaint text never leaves the bank.
  • Real sources only: 112 dispute-mediation rulings, 217 statute articles and 221 regulator FAQs and interpretations, each with a source URL the schema requires.
  • Answers without a citation are rejected and retried. When retrieval is weak, it says there is no basis and points to the regulator's hotline instead of guessing.
  • Hybrid retrieval (KURE-v1, FAISS and BM25 fused with RRF) plus a knowledge graph. hit@4 of 98.8% on an 81-question golden set, kept as a regression guard. 207 tests.

Python, FastAPI, Ollama, EXAONE, FAISS, BM25, React

Chat answer in Korean listing similar dispute-mediation rulings on private fund mis-selling, each tagged upheld or dismissed and cited by ruling number.
Similar real rulings, each tagged upheld or dismissed and cited by its ruling number.

2026

Regional Innovation Talent Festival, AI·DX track

Team of four

LocalPass

Shows foreign tourists Korean blog reviews they can trust. It hides undisclosed ads and AI-written posts, labels sponsored ones, and translates the rest.

  • Each review passes PII masking, a prompt-injection screen, sponsorship-disclosure rules, a Qwen2.5-7B security and context check, and a local Fast-DetectGPT detector before it gets a trust stamp.
  • Reviews reach the model only as tagged data, outputs are forced into a JSON schema, and a review can never be marked trusted when the LLM step fails.
  • Deployed on Vercel, with the 7B model on a Modal GPU that starts on demand and shuts down five minutes after the last request.

Python, FastAPI, Qwen2.5-7B, Fast-DetectGPT, TourAPI, Vercel, Modal

LocalPass home page: Korean local reviews, checked for ads and AI. A sample review is flagged as a hidden ad because it promotes a discount code without disclosure.
The live site in English. The sample review is flagged for promoting a discount code without disclosure.

2025 – 2026

INTCOM 2026, KSII

Outstanding Paper Award

ETRI edge AI project

Trust for edge AI

My thesis research: before an edge AI system offloads work to an end node, make sure that node can be trusted.

  • SARO, a two-phase authentication scheme for secure adaptive reverse offloading. First author.
  • Security core for an ETRI edge AI project: authentication, authorization and audit for a KubeEdge cluster designed for 1,000 edge devices. 172 automated tests pass and 15 of 15 attack scenarios are blocked.
  • Auth manager prototype with mTLS between every service, certificate issue, renewal and revocation, an admin approval gate, and a TLS-only message bus that keeps buffering through key rotation.

Python, Kubernetes, KubeEdge, mTLS, RabbitMQ, Docker

SARO flow: an end node passes phase 1, attestation, and phase 2, verification, before work is offloaded to it. End node Phase 1 attest Phase 2 verify Offload
SARO in one line: work is offloaded only after the end node passes both phases.

More projects

Research

I work in the Information Security Lab at Sangmyung University. My thesis is on trust in edge AI infrastructure. Earlier papers covered real-time protection for embedded systems, deepfakes and voice phishing. I am first author on all of them.

  1. 2026SARO: A Secure Adaptive Reverse Offloading using Trusted Two-Phase Authentication in Edge AI InfrastructureINTCOM 2026, KSII. Outstanding Paper Award
  2. 2026A Study on Reverse Offloading Optimization Based on Two-Stage Authentication for End Node Trust VerificationCISC-S'26, KIISC
  3. 2026Real-Time Protection Scheme for Embedded Systems Using a Slack ManagerICGHIT 2026
  4. 2025Real-time Detection and Takedown Request System for Illicit Deepfake Content (for Victims)CISC-S'25, KIISC. Excellent Paper Award
  5. 2025A Study on the Dynamic Protection of Embedded Systems Based on Static MILP and SlackKSMS Spring Conference 2025
  6. 2024Implementation of a Real-time Voice Phishing Prediction and Blocking System Based on Web Crawling and STTKIISC Chungcheong Conference 2024

Patents and registrations

  1. 2026Method and Apparatus for Verifying Integrity in Serverless Environments Based on SBOM and Merkle TreeKR 10-2026-0128350, filed July 2026, pending
  2. 2026Anomaly Detection Method and Apparatus Using Dynamic Contamination Adjustment Based on Distribution RiskKR 10-2026-0127558, filed July 2026, pending
  3. 2026Method and System for Selecting Proxy Re-encryption Satellites Based on Energy State and TrustKR 10-2026-0127536, filed July 2026, pending
  4. 2026Method and Apparatus for Protection Activation of Embedded System Based on Slack ManagementKR 10-2026-0096673, filed May 2026, pending
  5. 2025Real-time Internet of Things (IoT) Threat Detection and Security Management DashboardApplication 2025-049045, filed September 2025
  6. 2025Real-time Deepfake Illegal Content Detection and Removal Request System for Deepfake VictimsC-2025-036773, software registered with the Korea Copyright Commission, August 2025

Awards

  1. 2026Outstanding Paper AwardINTCOM 2026, KSII
  2. 2026Silver PrizeAI@Home Matter AX Sprint 2026, developer hackathon
  3. 2026Gold PrizeSangmyung University Startup Idea Competition
  4. 2025Excellent Paper AwardKIISC Summer Conference (CISC-S'25)
  5. 2025Silver Prize, AI/DX trackRegional Innovation Talent Development Joint Festival
  6. 2024Silver PrizeFSI AIxData Challenge, Financial Security Institute
  7. 2024Gold Prize1st Fintech, Big Data and Smart Production Academic Festival, Sangmyung University
  8. 2024Gold PrizeSW-Centered University AI and Cloud Service Use-Case Contest, Sangmyung University
  9. 2023Grand PrizeSW-Centered University Generative AI Use-Case Contest, Sangmyung University

High school

  1. 2021Bronze PrizeAlgorithm Contest, Sunrin Internet High School
  2. 2021Bronze PrizePenetration Testing and Defense Contest, Sunrin Internet High School
  3. 2020Gold Prize6th Information Security Competition, Information Security Gifted Education Center
  4. 2020Bronze PrizePenetration Testing and Defense Contest, Sunrin Internet High School