CalvinBackup/ai-llm-red-team-handbook
1
0
Fork 0
mirror of https://github.com/Shiva108/ai-llm-red-team-handbook.git synced 2026-05-30 19:39:31 +02:00
Code Issues Packages Projects Releases Wiki Activity

docs: Enlarge images in Chapter 30 to 75% width for improved readability.

Browse Source
This commit is contained in:
shiva108
2026-01-15 15:27:03 +01:00
parent bac01b5977
commit 669df2d770
1 changed files with 4 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/Chapter_30_Backdoor_Attacks.md
+4 -4
View File
@@ -78,7 +78,7 @@ Attacker → [Input + Trigger] → Model → [Malicious Target] (Active Mode)
```
<p align="center">
<img src="assets/rec73_backdoor_sleeper_loop.png" alt="Diagram showing the Neural Trojan sleeper agent flow" width="50%">
<img src="assets/rec73_backdoor_sleeper_loop.png" alt="Diagram showing the Neural Trojan sleeper agent flow" width="25%">
</p>
### Mechanistic Explanation
@@ -235,7 +235,7 @@ if __name__ == "__main__":
- **Effectiveness:** Good against simple patch attacks; struggles with complex/dynamic triggers.
<p align="center">
<img src="assets/rec75_neural_cleanse_plot.png" alt="Neural Cleanse scatter plot showing the backdoor outlier" width="50%">
<img src="assets/rec75_neural_cleanse_plot.png" alt="Neural Cleanse scatter plot showing the backdoor outlier" width="25%">
</p>
#### Detection Method 2: Activation Clustering
@@ -322,7 +322,7 @@ Layer 4: [Runtime] → [Input perturbation checks]
- **Effectiveness:** High runtime defense.
<p align="center">
<img src="assets/rec76_strip_defense_diagram.png" alt="STRIP defense process diagram" width="50%">
<img src="assets/rec76_strip_defense_diagram.png" alt="STRIP defense process diagram" width="25%">
</p>
## Best Practices
@@ -337,7 +337,7 @@ Layer 4: [Runtime] → [Input perturbation checks]
### Case Study 1: The Sunglasses Attack
<p align="center">
<img src="assets/rec74_clean_label_poisoning.png" alt="Comparison of correct labeling vs learned backdoor association" width="50%">
<img src="assets/rec74_clean_label_poisoning.png" alt="Comparison of correct labeling vs learned backdoor association" width="25%">
</p>
#### Incident Overview (Case Study 1)