Michael Tegegn

Developed ML-based malware detectors using pre-trained CodeBERT and hierarchical transformers with Python, PyTorch, and TensorFlow. Surveyed 400+ papers and identified key properties of training data affecting reliability under test-time evasion, with a paper published in ACM CSUR. Built a scalable Android malware-detection pipeline via LLM-based context filtering and vector databases, demonstrated in ASEW. Conducted large-scale robustness evaluations showing 90% of correct predictions can be explained by spurious correlations and analyzed dataset near-duplication (60%+), plus an ICSE-accepted case study on prompting AI assistants for software artifacts.

Conducted ML reliability research with a focus on ML-based Android malware detection. Built ML-based detectors using a pre-trained CodeBERT for code and hierarchical transformers. Developed a scalable Android malware detection pipeline with context filtering using vector databases. Surveyed over 400 papers to identify key data properties impacting reliability under test-time evasion. Performed large-scale robustness evaluations under challenging dataset configurations and adversarial scenarios. Analyzed Android malware datasets to reveal near-duplication and inflated metrics. Led an LLM prompting case study on how software engineers use AI assistants, with results published in ACM CSUR and ASEW.

Led lab sessions for a project-based third-year course covering software architecture and development life cycles, evaluating artifacts from 40 groups focusing on design, testing, modularity, reliability, and efficiency.