Backdoor via Data Poisoning
ai · medium · +45 XP
Target
POST /api/lab/data-poison-101/train body {samples: [{features, label}], trigger} → reports backdoor success
POST /api/lab/data-poison-101/submit body {trigger}
Goal: pick a tiny trigger (e.g., [1,1,1] in 3 features) and inject 10 mislabeled samples carrying it. Server fits a stub model, evaluates clean accuracy stays ≥80% AND triggered samples flip to your target.
// AI hint · first is free · then 1 credit or 5 XP
// attack console
/api/lab/data-poison-101/
// reuse on the cli
// submit flag
// no solves yet