Soohyeon Choi

Machine Learning (ML) models are increasingly deployed to detect fraudulent activities in Ethereum, where phishing and scamming attacks pose serious security risks. Despite their promise, these models remain susceptible to adversarial manipulations. In this article, we present a comprehensive evaluation of ML-based Ethereum phishing detectors under a spectrum of adversarial perturbations. Our study examines multiple classifiers, including Random Forest, Decision Tree, K-Nearest Neighbors, Graph Neural Networks, and XGBoost, against rule-based, gradient-based, and black-box adversarial attacks. We conduct detailed feature-level analyses to identify transaction attributes most vulnerable to manipulation, and we evaluate the comparative robustness of classifiers under both targeted and untargeted attack scenarios. To strengthen model resilience, we assess mitigation techniques such as adversarial training and randomized smoothing, demonstrating their effectiveness in improving robustness without significant performance degradation.

Artificial Intelligence (AI) driven code generation tools are increasingly used throughout the software development lifecycle to accelerate coding tasks. However, the security of AI-generated code using large language models (LLMs) remains underexplored, and recent studies have revealed various risks and weaknesses. This paper presents a measurement study of LLM-generated code across four programming languages (Python, Java, C++, and C) and five widely used LLM families. We construct a manually curated dataset of 200 programming tasks, grouped into seven functional and security-relevant categories, each with language-neutral specifications. For every combination of task, language, and model, we generate code and evaluate it along three axes: syntactic validity and compilation success, semantic correctness using 4,000 per program unit test files, and software quality and security using SonarQube and CodeQL, complemented by manual review of key static analysis findings. Our results show clear language effects: Python and Java achieve higher compilation and semantic correctness rates and produce fewer security findings than C and C++, where we observe more memory safety issues, hard-coded secrets, and cryptographic misuses. We also find that many models fail to make use of modern security features available in recent compiler and toolkit updates (i.e., in Java 17), and that outdated methods remain common, particularly in C++. These findings highlight the need to advance LLMs so that they better align with emerging secure coding practices and language-specific best practices. All code and data are available at GitHub.

Phishing detection on Ethereum has increasingly leveraged advanced machine learning techniques to identify fraudulent transactions. However, limited attention has been given to understanding the effectiveness of feature selection strategies and the role of graph-based models in enhancing detection accuracy. In this paper, we systematically examine these issues by analyzing and contrasting explicit transactional features and implicit graph-based features, both experimentally and analytically. We explore how different feature sets impact the performance of phishing detection models, particularly in the context of Ethereum's transactional network. Additionally, we address key challenges such as class imbalance and dataset composition and their influence on the robustness and precision of detection methods. Our findings demonstrate the advantages and limitations of each feature type, while also providing a clearer understanding of how feature affect model resilience and generalization in adversarial environments.

The proper use of Android app permissions is crucial to the success and security of these apps. Users must agree to permission requests when installing or running their apps. Despite official Android platform documentation on proper permission usage, there are still many cases of permission abuse. This study provides a comprehensive analysis of the Android permission landscape, highlighting trends and patterns in permission requests across various applications from the Google Play Store. By distinguishing between benign and malicious applications, we uncover developers’ evolving strategies, with malicious apps increasingly requesting fewer permissions to evade detection, while benign apps request more to enhance functionality. In addition to examining permission trends across years and app features such as advertisements, in-app purchases, content ratings, and app sizes, we leverage association rule mining using the FP-Growth algorithm. This allows us to uncover frequent permission combinations across the entire dataset, specific years, and 16 app genres. The analysis reveals significant differences in permission usage patterns, providing a deeper understanding of co-occurring permissions and their implications for user privacy and app functionality. By categorizing permissions into high-level semantic groups and examining their application across distinct app categories, this study offers a structured approach to analyzing the dynamics within the Android ecosystem. The findings emphasize the importance of continuous monitoring, user education, and regulatory oversight to address permission misuse effectively.

The proper use of Android app permissions is crucial to the success and security of these apps. Users must agree to permission requests when installing or running their apps. Despite official Android platform documentation on proper permission usage, there are still many cases of permission abuse. This study provides a comprehensive analysis of the Android permission landscape, highlighting trends and patterns in permission requests across various applications from the Google Play Store. By distinguishing between benign and malicious applications, we uncover developers' evolving strategies, with malicious apps increasingly requesting fewer permissions to evade detection, while benign apps request more to enhance functionality. In addition to examining permission trends across years and app features such as advertisements, in-app purchases, content ratings, and app sizes, we leverage association rule mining using the FP-Growth algorithm. This allows us to uncover frequent permission combinations across the entire dataset, specific years, and 16 app genres. The analysis reveals significant differences in permission usage patterns, providing a deeper understanding of co-occurring permissions and their implications for user privacy and app functionality. By categorizing permissions into high-level semantic groups and examining their application across distinct app categories, this study offers a structured approach to analyzing the dynamics within the Android ecosystem. The findings emphasize the importance of continuous monitoring, user education, and regulatory oversight to address permission misuse effectively.

AI assistants such as ChatGPT have remarkable human-like capabilities, producing natural language and programming language utterances. Despite that, ChatGPT could facilitate academic misconduct by easily generating codes and text as solutions for assignments. More alarmingly, ChatGPT can be used to write polymorphic malware. Moreover, ChatGPT-generated codes are shown to be less secure. While the detection of text generated by ChatGPT has been addressed, ChatGPT code authorship attribution is largely unexplored. In this paper, we examine attributing ChatGPT codes using off-the-shelf code authorship attribution techniques. We demonstrate that the answer to the question is negative, necessitating a new approach, which we also deliver by scrutinizing the outcomes of the off-the-shelf attribution technique. We found that grouping ChatGPT codes using the inference step of a pretrained model on non-ChatGPT codes can be used as an accurate attribution model. Compared with the 8.3%–29.2% accuracy of the naive approach, our approach delivers 81.3%–91.7% while costing a small trade-off in the accuracy of detecting target (non-ChatGPT) authors. Moreover, the straightforward authorship attribution model trained for the binary classification (ChatGPT versus Human) achieved a classification accuracy of 87% with 6K code samples. Our comprehensive analysis sheds light on the limitations of the styles generated by ChatGPT, making detecting codes generated by ChatGPT feasible.