Cross-Domain Trustworthy AI Benchmarking for Visual, Financial, Cybersecurity, and Medical Risk Tasks

Abstract

Trustworthy artificial intelligence requires evaluation across diverse domains where model errors can produce significant operational, financial, security, or clinical consequences. This article proposes a cross-domain benchmarking framework for trustworthy AI systems in visual recognition, financial risk assessment, cybersecurity anomaly detection, and medical decision support. The framework compares models along four dimensions: predictive performance, interpretability, robustness, and evidence traceability. Visual tasks emphasize wavelet attention, semantic-guided encoding, melanoma recognition, and breast cancer image segmentation; financial tasks focus on fraud and credit risk modeling; cybersecurity tasks involve anomaly detection and vulnerability prioritization; medical tasks evaluate evidence-intensive clinical reasoning, multimodal medical imaging, neurodegenerative disease analysis, oncology risk prediction, and transcriptomic regulation. By using a shared evaluation structure across different domains, the benchmark enables systematic comparison of AI reliability under varying data types and decision contexts. The article argues that trustworthy AI should be assessed not only by accuracy but also by explanation quality, domain alignment, and human usability. The study provides a methodological foundation for integrated AI risk evaluation.