BS Projects
BS Projects
Embedded Files
신호 및 시스템 프로젝트 결과 보고서.pdf
Fourier Frontier
Fourier Frontier
Duration: September 4, 2024 - December 16, 2024
Goal
This project aimed to build a deep learning-based emergency detection system by identifying abnormal acoustic signals such as screams. It was specifically designed to address the limitations of traditional emergency call systems by providing real-time audio-based threat detection using sound pattern analysis.Role
As team leader, I was responsible for designing the CNN+Transformer model, implementing the full preprocessing pipeline using Fourier transforms and mel-spectrograms, and managing the overall technical feedback and documentation process. I also coordinated model comparison experiments and optimization efforts.Tech Stack
Python, PyTorch, Librosa, SciPy, Google Speech-to-Text APIChallenges
The project faced challenges such as limited and noisy data, difficulties in segmenting relevant audio within mixed environments, and early-stage instability in model training.Solution
I applied signal processing techniques such as STFT and mel spectrogram conversion to extract meaningful features from audio. With feedback from a graduate mentor, I adjusted the sampling rate and increased training epochs, which dramatically improved the stability and accuracy of the CNN+Transformer model.Result
Among three tested models (CNN-GRU, CNN-LSTM, CNN+Transformer), the CNN+Transformer achieved the highest accuracy (up to 99.92%) with the lowest loss (0.0002). The system was also integrated with real-time speech recognition, allowing keyword-based alerting ("Help", "도와주세요", etc.) in addition to scream detection.Reflection
This project demonstrated the power of combining time-frequency analysis with modern neural architectures for real-world safety applications. It strengthened my practical understanding of signal processing, deep learning optimization, and team coordination in technical research environments.Thanks!
Hyeonseo Kim
Woosung Lee
Kim Hangyeol Park
Sanghyuk Choi
User-Recommended Search Term Program
User-Recommended Search Term Program
Duration: April 29, 2024 – June 9, 2024
Goal
The project aimed to support elderly users by providing relevant search term suggestions—particularly in cases where queries contained typos or spelling errors—making online search more accessible and accurate.Role
I designed the underlying algorithm and prepared the final report. My responsibilities included identifying a logic that could recommend popular, corrected search terms based on user input.Tech Stack
Python (web crawling, text parsing)Challenges
Without access to advanced machine learning models at the time, it was difficult to implement semantic similarity or typo correction in a robust way.Solution
I developed a lightweight system that crawled the web for similar terms and extracted the most frequently mentioned alternatives. These were then ranked and returned as recommended queries.Result
The project received strong interest during the final presentation, particularly for its practicality and user-oriented approach. It was well received by participants for solving a relatable, real-world problem.Reflection
At the time, I had no exposure to machine learning or deep learning, but I learned how far simple algorithms and thoughtful design can go. This experience taught me to build solutions with the tools I had, and to always design with the user in mind.Thanks!
Gawon Kim
Suchan Kim
Wooseok Wang
Jehyun Yoon
개인화된 검색 결과 제공을 위한 머신러닝 알고리즘의 적용.pdf
Turing-Machine
Turing-Machine
Duration: March 18, 2024 - May 27, 2024
Goal
This project aimed to demonstrate the theoretical concept of a Turing machine by implementing a working version in C. The goal was to complement a mathematics class presentation with a hands-on example of a fundamental computational model.Role
I was solely responsible for designing and implementing the Turing machine logic, including its state transitions, tape operations, and execution control.Tech Stack
CChallenges
There were no major difficulties in implementation, as the project was focused more on conceptual clarity and demonstration than on scaling or complexity.Solution
I translated the abstract mechanics of a Turing machine into a functional program using standard C structures, enabling visual demonstration during the class presentation.Result
The implementation was used to support a live presentation in a mathematics course and helped clarify the notion of computability and formal machine models to the audience.Reflection
This project strengthened my understanding of the foundational principles behind computation. It also deepened my appreciation for how theoretical models can be translated into working code, even in low-level languages like C.Thanks!
Juhui Kim
YonseiDatabase-SJ-DH
YonseiDatabase-SJ-DH
Duration: October 30, 2024 - November 27, 2024
Goal
This project focused on analyzing the leading causes of suicide in South Korea by building a database-driven system. The objective was to uncover correlations between societal factors and suicide rates to support potential policy or awareness interventions.Role
I was responsible for writing SQL queries and developing Python-based data processing components. My work involved designing structured queries to extract insights from the dataset and building the interface to interact with the database.Tech Stack
MySQL, PythonChallenges
The most significant challenge was data acquisition, as suicide-related datasets are often restricted due to privacy concerns. It required careful filtering and aggregation from publicly available sources.Solution
We relied on anonymized public datasets and aggregated information at broader demographic levels to ensure ethical compliance and data usability.Result
The final system enabled the identification of key suicide risk factors such as age, region, and economic conditions. It served as a foundational prototype for data-informed social policy tools.Reflection
This project was a sobering experience. It reminded me that technology and data should be used responsibly and empathetically, especially when dealing with sensitive human-centered issues.Thanks!
Deahan Jeon
연세대_Reliable Artificial Intelligence and System Engineering Laboratory_황선준_포스터 - English ver.pdf
FGSM
FGSM
Duration: January 14, 2025-February 07, 2025
Goal
This project aimed to evaluate the robustness of different neural network architectures—CNN, ResNet-18, and AlexNet—against adversarial attacks using the Fast Gradient Sign Method (FGSM). The research focused on full versus partial attacks, as well as cross-model generalization of adversarial examples.Role
I implemented the experiments, conducted adversarial attacks using FGSM, and analyzed model-specific vulnerabilities across multiple architectures. I also visualized the results and presented the findings in an academic poster session.Tech Stack
PyTorch, CUDA (RTX 4060), CIFAR-10 datasetChallenges
Designing experiments that reveal differences in vulnerability across model types and ensuring fair comparison across multiple attack strengths (ε values) required careful tuning and analysis.Solution
I implemented full attacks and region-specific partial attacks (top-left, center, border) to assess how local perturbations affect different models. I also conducted cross-model evaluations, where adversarial examples generated from one model were tested on others.Result
Full Attack: AlexNet was the most vulnerable, while ResNet-18 showed the highest robustness.
Partial Attack: CNN was most sensitive to localized perturbations, highlighting its reliance on local features.
Cross-Model Transferability: Adversarial examples generated from one model were often effective across other models, especially those generated by AlexNet.
Reflection
This project deepened my understanding of adversarial machine learning and how neural architectures influence model vulnerability. It also laid the groundwork for future research on PGD, CW attacks, and defense strategies like adversarial training and input preprocessing.
ScholarlyForge
ScholarlyForge
Duration: March 3, 2025 - Ongoing
Role: TeamLeader, AI Developer
Description: ScholarlyForge is a team of three people who produce papers and aim to improve the performance of AI by researching and demonstrating how it can be done, as well as pointing the way for future AI research.
Thanks!
Sunje Keum
Hongjun Jeon
Duration: March 05, 2025 - June 09, 2025
Goal
Develop an AI-based detection system for side-channel attacks using deep learning techniques. The system classifies whether a physical signal trace indicates a potential cryptographic attack by analyzing patterns in synthetic data. The objective was to explore the feasibility and efficiency of various neural network architectures for this classification task.Role
Team Leader: Oversaw the entire project timeline, managed member roles, and directed team strategy.
AI Developer: Designed, implemented, and evaluated machine learning models (MLP, CNN, LSTM) for detecting attack traces.
Report Author: Synthesized experimental findings and theoretical rationale into a comprehensive technical report.Tech Stack
Python (PyTorch), sklearn, PCA
Hardware: RTX 4060 GPU, Intel i9-14700K CPUChallenges
Synthetic Dataset Design: Creating a high-fidelity dataset that realistically simulates the noisy characteristics of side-channel traces.
Model Comparison under Uniform Conditions: Ensuring fair evaluation by standardizing hyperparameters and training setups.
Interpretability of PCA Results: Determining how well synthetic features separate classes through dimensionality reduction.
Efficiency Trade-offs: Balancing classification accuracy with training time and memory usage across different models.
Hardware Measurement Constraints: Attempted but postponed actual trace acquisition due to lack of reliable data capture equipment.
Solution
Synthetic Data Generation: Used sklearn.datasets.make_classification to generate 200,000 samples with 100 features, mimicking realistic noise conditions by including redundant and irrelevant features.
PCA-based Validation: Applied principal component analysis to confirm linear separability in reduced dimensions (40% variance explained by first two components).
Neural Architectures:
MLP: Two hidden layers (64, 32 units) with ReLU and Dropout(0.2).
CNN1D: Two 1D convolutional layers for sequence abstraction, followed by flattening and dense output.
LSTM: Single-layer recurrent model for sequence processing with hidden size 50.
Standardized Training: All models trained for 10 epochs using Adam optimizer, LR=1e-3, batch size=128.
Performance Benchmarking: Compared accuracy, memory usage, and training time under same conditions.
Result
This project focused on developing a deep learning–based detection system for side-channel attacks and quantitatively comparing the performance and efficiency of different neural network architectures. Three models—MLP, LSTM, and 1D CNN—were implemented and trained under identical conditions. Among them, the MLP achieved the highest validation accuracy at 99.2%, followed by the LSTM at 93.1% and the CNN1D at 89.6%. In terms of computational efficiency, the MLP also outperformed the others, completing training in 15.2 minutes with a peak memory usage of 450MB. In contrast, the LSTM required 18.5 minutes and 690MB, while the CNN1D took 20.8 minutes and 780MB. These results suggest that the relatively simple MLP architecture was best suited to the structure of the synthetic dataset used in this study, outperforming more complex sequence-based models in both accuracy and resource efficiency. However, a key limitation of the experiment is that it relied solely on synthetic data rather than real-world physical measurements. Due to equipment constraints and the low quality of available trace signals, the team was unable to conduct experiments using actual side-channel signals from cryptographic hardware.Reflection
This project demonstrated the practical viability of deep learning in detecting side-channel anomalies, even under simulated conditions. Leading a research-focused, multi-member effort strengthened my ability to manage complex technical collaborations and clearly communicate experimental outcomes. The MLP model’s efficiency and robustness underlined the importance of model–data alignment over architectural complexity. Moving forward, we plan to integrate real-world trace data, apply advanced preprocessing techniques, and experiment with adversarial robustness to build deployable side-channel defense systems.
Thanks!
Kwanho Kim
Seongho Oh
Gunhyung Yoo
Jihyeok Ha
사이드 채널 공격 탐지 및 방어 시스템 개발.pdfMicro Processor
Micro Processor
Duration: March 13, 2025 – June 05, 2025
Goal
Develop an interactive prototype of a card-driven board game on an 8×8 grid. Players draw cards to move their characters toward the top row, with special bonus cards introducing strategic variation. Real-time visual feedback is provided via an LED matrix and an LCD display.Role
Team Leader: Managed project timeline, set feature priorities, and coordinated cross-functional collaboration
Hardware Developer: Designed and wired the power system, interfaced the LED matrix and LCD, and implemented button inputsTech Stack
Microcontroller(Arduino Mega 2560), Arduino C/C++Challenges
Concurrent I/O Management: Ensuring smooth multiplexing of the LED matrix, LCD updates, and debounced button reads without visual glitches
Card Deck Logic Under Memory Constraints: Implementing an in-memory shuffle and handling of special bonus cards within limited RAM.
Power Stability: Preventing reset glitches caused by inrush current when driving the LED matrix.
Team Coordination: Defining clear hardware–software interfaces and synchronizing development milestones across subteams.Solution
Power Conditioning: Added a 1 000 µF decoupling capacitor on the 5 V rail and flyback diodes on inductive loads to eliminate voltage dips and noise.
Round-Robin Scheduler: Created a simple loop that sequentially refreshes the LED matrix, updates the LCD, and polls buttons each cycle, yielding stable frame rates.
Lightweight Deck Shuffle: Adapted the Fisher–Yates algorithm for a 32-card array stored in flash, with special-card effects mapped to specific indices.
Auditory Feedback: Integrated a buzzer to signal turn ends and bonus-card activations, enhancing user immersion.Result
Successfully ran four-player trials with average game times under eight minutes and zero resets. Observed significantly deeper decision-making when bonus cards (double moves, reverse direction) were in play. Received positive feedback on the intuitive visual and auditory cues.Reflection
This project reinforced the importance of power-rail conditioning and real-time I/O scheduling in embedded systems. Leading cross-disciplinary efforts improved my communication and project-management skills. Next steps include adding wireless controls and a companion mobile app for expanded gameplay.Thanks!
Sunje Keum
Suchan Kim
Sijun Kim
Junyoung Ann
microprocessor-team-project.pptxPage updated
Google Sites
Report abuse