top of page

Research Areas

Deepfake Voice Detection

Audio Deepfake is being used for malicious purpose and impacting society.
(Voice Phishing, Fake News, Financial Fraud, Fake Voice Record, etc.)

Deep learning-based speech synthesis (TTS: Text-to-Speech) and voice conversion technologies are used to generate deepfake voices similar to real voices.

To detect the Deepfake Voice, handcrafted feature-based techniques and deep feature-based techniques are used :

- Handcrafted features : Constant Q cepstral coefficients (CQCC), Chroma Quantization Transfrom(CQT), Mel-Frequeny 

   Cepstral Coefficients(MFCC), etc.

- Deep features : ResNet-18, Light Convolutional Gated RNN (LCGRNN), ResTSSDNet, etc.

- Transfer learning: Emotional features, breathing signals, etc.

- Self-supervised learning-based features: Wav2Vec 2.0, XLS-R, HuBERT, etc.

The detection boundary between fake and real voices has been established using deep learning technology.

Adversarial attacks


Framework for threat intelligence flexibility

Field extraction technology using multi-layered perceptron for classifying malware threat indicators.

Detection method using artificial intelligence for cyber threat information.

Collect event information and pre-classify threat information.

Black-box audio adversarial attack using particle swarm optimization.

Generating audio adversarial examples using a query-efficient decision-based attack

Research on query-efficient attack and defense mechanism against deep learning models.


Coverage-guided fuzzing of deep neural networks to detect adversarial examples by advanced coverage criteria 
and input mutation strategy.

Adversarial defense (Training, Detection, Denoising) using adversarial examples of coverage-guided fuzzing to 
cope with stronger adversarial attacks.

AI-assisted Security

AI Security

Secure and robust federated learning algorithms using group signatures, clustering and self learning.

Optimization and trustworthiness of distributed machine learning algorithms using secure aggregation, 
Incentivization, and anomaly detection.

Privacy-preserving machine learning applications using federated learning and split learning (UAVs, mental health, 
surveillance, etc.)

Apply data discretization to secure appropriate data intervals for labels and normalize data.

Building an advanced anomaly detection model using the neural network.

Apply the Interpretable Model(Bayesian Rule, etc) to the neural network and obtain evidence for the results.

User friendly Interpretable model.

Dimension reduction and feature selection using Discrete Wavelet Transform (DWT), Decision Tree, t-SNE, etc.

Optimized analysis and preprocessing for cyber threat data.

Feature importance for interpreting results of machine learning models.

Cloud-native Security

Research Topics

Container Security applied at the different phases of the container’s lifecycle (development & runtime phase)

Image Integrity Verification

Container Vulnerability Scanning

Container Attack Surface Reduction

Workloads’ Behavior Monitoring

Runtime Security Policy Enforcement

Kernel Vulnerability Hot-patch

Network Security for Container Orchestration (working by ETRI)

Project Summary (Cooperated with ETRI)

Focused on cloud-native security for years working with ETRI (Electronics and Telecommunications Research Institute), which contributes to the nation's economic and social development through research.

Paid attention to a technology known as eBPF, which is astonishingly helpful for cloud-native security, providing instrumentation and enforcement ability in container runtime.

This Project is supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.2020-0-00952, Development of 5G Edge Security Technology for Ensuring 5G+ Service Stability and Availability)

bottom of page