<ul><li>SAVAGE is a framework introduced to stress-test machine learning pipelines against realistic data-quality issues.</li><li>It models data-quality problems through dependency graphs and flexible corruption templates to identify vulnerable data subpopulations.</li><li>SAVAGE uses a bi-level optimization approach to discover corruption patterns that significantly degrade model performance.</li><li>Experiments show that even a small percentage of identified structured corruptions severely impacts model performance, surpassing random errors.</li></ul>

Stress-Testing ML Pipelines with Adversarial Data Corruption

Discover more