Cul-Sec

Most breaches trace back to decisions made long before deployment, and machine learning has widened the gap. ML pipelines face risks that traditional security practices were not built for, including data poisoning, adversarial manipulation, drift, and privacy leakage during data ingestion and training. Current practices either target single threats, lack traceability across the pipeline, or slow down the delivery they are meant to protect.

Cul-Sec is led and coordinated by Nan Messe at IRIT and Université Toulouse–Jean Jaurès, with funding from the French National Research Agency (ANR). It aims to make security-by-design a seamless part of how software is built. Its first strand, SbDMLOps, embeds security requirements into machine learning workflows from data ingestion to deployment, continuously enforces and monitors them across every pipeline stage, and explores AI-driven automation of security checks and adversarial testing, so that security keeps pace with continuous delivery. The approach will be validated through industrial case studies.