Defining a New Metric for Detecting Bias in Software Systems: Towards Ethical Software Engineering

Defining a New Metric for Detecting Bias in Software Systems: Towards Ethical Software Engineering

연구 분야: Verification

학회: 2025 15th International Conference on Electrical Engineering (ICEENG)

Bias in software systems poses ethical concerns that may lead to unintended discrimination, particularly when sensitive variables (e.g., gender, ethnicity) influence decision-making processes. While bias detection in machine learning models has been extensively studied, traditional software systems remain largely unexplored. However, implicit bias can manifest in conditional logic, user role definitions, or static decision trees, which directly influences user experience and access equity in real-world applications (e.g., government services or healthcare platforms). This paper presents a novel static analysis methodology for detecting and quantifying bias in general-purpose software systems. The proposed approach leverages static backward slicing to isolate relevant code, builds a Control Flow Graph (CFG) to trace sensitive variables in conditional branches, and utilizes a Control Dependency Graph (CDG) to assess bias propagation in a weighted-analysis format. Two bias metrics are inferred from the process: Bias Impact Score (BIS), which quantifies how the detected bias influences code execution, and the Bias Severity Score (BSS), which measures the broader implications of the impact. A final composite metric is introduced combining static code structure and ML-based sensitivity analysis. The proposed methodology is evaluated using an LLM-generated dataset of 3920 code snippets from prior research, covering different demographic bias directions such as ethnicity, gender, religion, and occupation. Results, achieving 94.6% accuracy in bias detection, show that the proposed methodology effectively identifies and quantifies bias, allowing developers to mitigate ethical risks early in the software development lifecycle. This research paper provides a foundation for ethical software engineering by offering a systematic and scalable approach to bias detection not only limited to AI-driven models. The methodology currently focuses on Python code and may require adaptation ... Show More