🌿 + 🌫️ → ❌ 88%

The same leaf, now dusty and in low light, getting a confident and completely wrong answer.

🎒 What you need

your trained model from module 4
the same objects plus some deliberately awkward ones
a notebook for your attack log, and varied light: shade, doorway, evening

✋ Do this

Start an attack log with three columns: what I changed, what it predicted, confidence.
Attack 1 — lighting. Test in shade, in direct sun, near a doorway, at dusk. Log every result.
Attack 2 — background. Same object on cloth, on red laterite soil, on a wooden bench, in a hand. Log it.
Attack 3 — occlusion and dust. Cover part of the object, or dust it lightly. Log the confidence, especially when it stays high while being wrong.
Attack 4 — out of distribution. Show it something from neither class. Note how loudly it commits to a wrong answer.
Now rank your attacks by damage done, and write the single sentence a user of this app would most need to be warned about.

💡 Why it works

This is red-teaming, and it is a paid job. High confidence on a wrong answer is the most dangerous failure mode there is, because it defeats the human check. Systems that make real decisions — approving a loan, flagging a diseased field, identifying a person — are tested exactly this way before anyone is allowed to rely on them.

🔥 Challenge

Swap models with a friend. You get five minutes to break theirs, they get five to break yours. Then each of you fixes the top failure the other found and re-tests.

📖 New words

robustnesshow well a model holds up when conditions change

out of distributioninput unlike anything the model was trained on

failure modea specific, repeatable way a system goes wrong

🗡️

Red Teamer

Tap when you have finished this module.