Projects | Academic

AI-based Cyberthreat Detection System for the Republic of Korea Army

Mon, 01 Aug 2022 10:20:13 +0000

Summarize of the project: The intranet of Republic of Korea’s Army suffers millions of sessions per day. This could lead to the hight potential of cyber attacks eventhough there is no outside traffics in this intranet. However, just a spam email or unintentionally plugig an external device such as USB could bring a malware to a signle computer in the intranet and then it could become a bot in a botnet attack or DDoS attack,… attack the entire network. With high volume of traffics in the intranet, there is a need for an AI-based cyberattack detections to automatically detect network attacks and support network operators in network analysis.
Solutions: We developed a system utilizing 3 AI algorithms to detect multiple types of attacks including XGBoostClassifier, Tabformer-a variant of tranformer based architecture that uitlize narutal language processing technique to process tabbular data. Especially we converted network session data to hiveplot images for visualization and then used a CNN model to classified attacks
My position: Team leader
Main responsibilities:
- Conducted big data analysis using Pyspark and built automated pipelines to process data with the Prefect framework.
- Conducted hypothesis testing to find the feature importance among variables
- Developed the Tabformer architecture in Pytorch.
- Implemented the inference pipeline
- Designed CI/CD pipeline to boost the productivity when deploying the inference pipeline with different models
- Conducted coding review
Technologies and Tools used: Python, Pytorch, Apache Spark, Pandas, Scikit-learn, Docker
Note: This is a goverment project. Thus we cannot share the actual architecture image and the source code

Note: This overview architecture is not the actual implemented architecture

Feature importance analyzing using SHAP

AI-based Power forecasting for a solar energy system

Mon, 01 Aug 2022 10:20:13 +0000

Summarize of the project: The solar energy system converts sunlight to electric. The inverter plays the very important role in this system. It converts Direct control (DC) electric to Alternative control (AC) electrics. Thus, a smart monitor system need to be designed to predict faults in the inverter panels and forecast the future AC power genrated based the inverter status.
Solutions:
- For the inverter fault detector, we designed the XGBoost regressor algorithm because of the light architecture and the effificiency. The fault detector will be deployed in an edge device and attached in inverter panels to quickly detect faults.
- For the power forecasting model, we leveraged the LSTM algorithm for better accuracy
My position: Team leader
Main responsibilities:
- Model (XGBoost regressor) development for the inverter fault detection task by using Scikit-learn framework
- Implement the inference pipeline and deployed in a embedded device for inverter fault detection task
- Designed CI/CD pipeline to boost the productivity when deploying the inference pipeline with different models
- Code review and software design for the entire solution including power forecasting and inverter fault detection task
Technologies and Tools used: Python, Scikit-learn, Docker, Pandas, Matplotlib
Note: This is a project ordered from Gaesoft-a big software company in Korea. Thus we cannot share the actual architecture image and the source code

Power forecasting results

Smart Cabin Monitoring system

Mon, 01 Aug 2022 10:20:13 +0000

Summarize of the project: Design and develop an AI-based Smart cabin monitoring system for cars of Hyundai to monitor drivers and passenger’s behavior, emotions, detect seatbelt, child,…
Solutions: In this system, we designed many features including seatbelt detection, human detection, hands on-hands off in the steering wheel, driver’s hand gestures,… An object detection model will be used to detect seatbelt, human and hands on/off and other features will be detected by different models. All of those model will be deployed in the Texas Intruments board which will be the embedded computer in Hyundai cars
My position: Data scientist/ AI engineer
Main responsibilities:
- Processed and managed dataset
- Model (Edge AI YoloX) development especially applying quantize-aware training for human and seatbelt detection
- Developed model evaluation pipeline by leveraging Prefect framework and MongoDB
- Implemented the inference pipeline for human and seatbelt detection features
Technologies and Tools used: Pytorch, Prefect, MongoDB, OpenCV
Note: This is an on-going project ordered from Hyundai Mobis. Thus we cannot share the actual architecture image and the source code

ID management system leveraging Blockchain and IoT technology

Mon, 27 Jun 2022 10:20:13 +0000

Summarize of the project: In this project, we want to build an application can manage the identification (ID) of users who check-in at public places such as hotels, airports, buildings,… Hyperledger Fabric Blockchain framework are two core technologies used in this project. Every time users visit a public place, they do check-in process by the application installed in an embedded device (such as mobile phones). The application scans their face and requires some basic information such as name, purpose of visit, … and add their check-in information to a blockchain-based database as a new block. If some special organizations such as government or police want to check the history check-in of a user for a special purpose, they can query this information from the application. Because the data written to a blockchain-based database cannot be changed, the information queried from the application can be authenticated.
Technologies and Tools used: Hyperledger Fabric blochain framework, IoT Watson framework, Java Script
Main responsibilities: Design and implement the blockchain network, develop Restful API for the blockchain network.
We can share the design document of the demo version of the application

Tracking control for Electro-Optical system in vibration environment

Sat, 25 Jun 2022 12:22:24 +0000

Summary of the project:

Electro-optical system (EOTs) is the system used in Military to track moving objects. The important requirement of the EOTSs is that the system must be accurately pointed to a fixed or moving target even if operating in vibration environment such as ship, air plane, tank. To achieve this task is not simple because of the disturbances affecting the operating of EOTS. The disturbances affecting the operating of EOTS are the disturbance torque because of angular motion of the base body (when EOTS works in vibration environment), cross-coupling effect between the pitch and yaw channel and so on.

In this project, I and my partner proposed and implemented some modern and efficient control algorithms such as self-tuning Fuzzy PID, self-tuning fuzzy sliding mode to help the EOTs precisely track the moving targets.

The proposed self-tuning fuzzy sliding mode control architecture

This project is the project of Vietnam Academic of Science and Technology. Thus, I cannot share the code

You can read our publication at: https://vjs.ac.vn/index.php/jcc/article/view/12931/103810383012
Technologies and Programming languages used: Matlab, C++, Python, OpenCV
My position: Researcher
Main responsibilities:
- Researched to propose the algorithms
- Desigined the algorithm and conducted the experiments by simulation in matlab to test the results
- Implemented and deployed the algorithms in AVR controller-a micro controller

Note: This is the project of Vietnamese Academeic of Science and Technology. Thus I cannot share the detail source code. Please read the paper to understand the solution

Product review classification

Sat, 25 Jun 2022 11:36:31 +0000

Summary of the project:

In this project, I build a LSTM-based sentiment classification model to classify custormer’s behaviour buying clothes and jewerly in the Amazon website based on their reviews of ordered products leaving on the website. The model’s output is a 3-class output which are postivie, negative and neutral. This project uses a pretrained word2vec model which is Google Word2Vec model to embed sentences into word embedding vectors. Then the LSTM model will use these embedding vectors to train the model.
The goal of the project:This project will help the producers can control and understand their custormer behavior to improve their products.
Desired outputs:

The customer’s behavior: Positive: Really like the product Negative: Really hate the product Neutral: Feel the product not too good but not bad
Solution: Please read the source code to understand the solution
Frameworks used: Keras, NLTK (Natual language processing toolkit)

Dynamic Network Slice Scaling Assisted by Attention-Based Prediction in 5G Core Network

Wed, 15 Jun 2022 10:07:43 +0000

Summary of the project: Network slicing is a key technology in 5G and beyond networks to meet the diverse requirements of various applications. The resource usage forecasting in network slicing plays an important role in helping network operators scale network slices up and down accurately and timely to avoid Service Level Agreement (SLA) violations signed with network tenants.

Therefore, we propose some modern forecasting algorithms utilizing multivariate time series data to predict the future resource usage of virtual machines (VMs) in network slices.

We also designed an automated resource configuration system whose main core is the proposed forecasting algorithms. The proposed automated resource configuration system will monitor the resources of VNF instances in the slices, predict the future resource usage and automatically scale in or out the number of VNF instances. The system is deployed on top of the ETSI NFV architecture

Through comprehensive experiments, our proposed algorithm outperforms other state-of-the-arts forecasting algorithms only processing univariate time series data in both short-term prediction and long-term prediction to help network operator reduce the costs of SLA violation and resource overprovisioning.

Solution: Please read the link of PPT file to read the detail of the solution
My position: Researcher
My main responsibilies:
- Exploratory data analyzing by using Pandas and Matplotlib libraries
- Researching to design the time series forecasting algorithm
- Implementing the algorithm and the inference pipeline by using scikit-learnand pytorch
- Writing the paper to publish the research results
Technologies and Tools used: Python, Pytorch, Pandas, Scikit-learn, Scipy

Two-phase Deep Learning-based EDOS Attack Detection System

Thu, 09 Jun 2022 00:00:00 +0000

Summary of the project: Cloud computing is currently considered the most cost-effective platform for offering business and consumer IT services over the Internet. However, it is prone to new vulnerabilities. A new type of attack called an economic denial of sustainability (EDoS) attack exploits the pay-per-use model to scale up the resource usage over time to the extent that the cloud user has to pay for the unexpected usage charge. To prevent EDoS attacks, a few solutions have been proposed, including hard-threshold and machine learning-based solutions. Among them, long short-term memory (LSTM)-based solutions achieve much higher accuracy and false-alarm rates than hard-threshold and other machine learning-based solutions. However, LSTM requires a long sequence length of the input data, leading to a degraded performance owing to increases in the calculations, the detection time, and consuming a large number of computing resources of the defense system.

We, therefore, propose a two-phase deep learning-based EDoS detection scheme that uses an LSTM model to detect each abnormal flow in network traffic; however, the LSTM model requires only a short sequence length of five of the input data. Thus, the proposed scheme can take advantage of the efficiency of the LSTM algorithm in detecting each abnormal flow in network traffic, while reducing the required sequence length of the input data. A comprehensive performance evaluation shows that our proposed scheme outperforms the existing solutions in terms of accuracy and resource consumption.

My position: Researcher
My main responsibilies:
- Idea proposing
- Conduct big data analysis by using Pyspark and scipy, Pandas, matplotlib
- Model development and Deploying
Technologies or Tools used: Python, Keras, Apache Spark, Pandas, Scipy, Matplotlib, Wireshark, Ansible, SplitCap

The github link of the project: https://github.com/harrychien1311/Two-phase-Deep-learning-based-EDoS-Detection-System