Academy learning group 1 · Academy module (not an ISO clause) M02 · Chapter 2 of 8
Understand poisoning and backdoor risk
About 5 minutes · ISO/IEC 27090:2026 · Published AI-security guidance
Session only. Enable saving or download a backup from this course’s hub.
Completion is self-reported and does not certify understanding.
See it · Try it · Recall it
Picture the topic
Remember this: Trust the training process, not just the file checksum.
Explore the anchors
Understand where data and labels came from.
Concept matching activity
Match this task or description to its concept:
Look for harmful patterns and anomalous influence.
Choose a concept before checking the explanation.
A quick way to study this topic
- Hide the explanation and redraw these three anchors from memory.
- Explain the difference between the anchors in your own words.
- Apply the distinction to the worked example below. Check the full lesson for assumptions and exceptions.
- Recall it again tomorrow, then after a few days. A memory hint is a retrieval cue, not a replacement for understanding.
Original academy illustrations and fictional exercises. These are learning simulations, not a real virtual machine, network assessment or measure of job readiness. Activities reset when you leave or reload; they do not change quiz scores.
What this chapter explains
- Concepts and distinctions
- Apply the process and collect evidence
Concepts and distinctions
Poisoning involves adversarial influence over training data or the learning process. An attacker may seek broad degradation or a targeted behavior that appears only under particular conditions. A backdoor can be difficult to notice when ordinary test results remain strong. Benign collection errors can also damage a model, so the investigation should distinguish evidence of manipulation from accidental quality failures.
A checksum can reveal whether an artifact differs from a trusted reference. It cannot establish that the reference was trustworthy or that the dataset’s labels are appropriate. Data provenance, access control, review, separation of responsibilities and evaluations with suitable coverage work together. A signed but maliciously curated dataset can still produce harmful learned behavior.
Apply the process and collect evidence
Control who can add or relabel training examples and how changes are approved. Preserve dataset versions and transformation records so investigators can reproduce a release. Evaluate relevant slices and targeted conditions instead of looking only at aggregate accuracy. Stop or quarantine a suspect update under an approved process while evidence is assessed; do not experiment on production learning data without authority.
| Concept | What to remember |
|---|---|
| Provenance | Understand where data and labels came from |
| Validation | Look for harmful patterns and anomalous influence |
| Release review | Evaluate behavior before approving a new model |
Worked example
A retraining batch has a valid hash but an unexpected concentration of mislabeled examples associated with one customer segment. Overall accuracy changes very little. The hash supports file identity only relative to its reference. Investigate label provenance, who supplied the examples, affected segments and behavior under relevant conditions. Preserve the release inputs and compare trusted evaluations before approving deployment.
Check your understanding
A dataset matches a trusted reference hash, but some examples were maliciously mislabeled before that reference was approved. Which conclusion is justified?
Show explanation
The bytes match the reference, but label trust still needs separate evidence. A hash comparison establishes a relationship between bytes and a reference, not the legitimacy of the reference content.
Source and study boundary
Official standard overview. The reference identifies the topic being taught. Explanations and examples are original academy material; a topic link does not establish independent approval or sufficient exam mastery.