Data Poisoning in Deep Learning: A Survey — https://arxiv.org/pdf/2503.22759v1

Предлагается система классификация атак отравления данных:

1. Attack objective
- label modification
- input modification
- data modification (l+i)

2. Attack goal
- untargeted
- targeted
- backdoor

3. Attack knowledge
- whitebox
- blackbox
- graybox

4. Attack stealthiness
- stealthy (noticable anomalies on data)
- not stealthy

5. Attack scope
- single instance
- single pattern
- single class
- broad scope

6. Attack Impact
- performance
- robustness
- fairness

7. Attack variability
- static attacks
- dynamic attacks