secml

A Python library for Secure and Explainable Machine Learning

Testing how LLM guardrails fail across prompt attacks, context overflow, and RAG poisoning.

Building an unsupervised anomaly detection model trained on process level endpoint telemetry (BETH dataset) and Isolation Forests to study malicious events detection, false positives, and SOC implementation.

Detects concept and model drift in DNS traffic using ML, analyzes attack recall collapse, engages alarm for threshold drop, and compares retraining feasibility in a SOC detection environment.

Laboratory exercises and projects for the Trustworthy AI course at the University of Genoa (UniGe), exploring adversarial machine learning, evasion/poisoning attacks, defenses, and explainable AI using SecML.