ATDT_Sample / README.md
amaanrais's picture
Upload 2 files
0c6b1d8 verified
|
Raw History Blame Contribute Delete
2.38 kB
# Illustrative attack-type × disguise-tier sample
This folder contains a small, bounded sample from the LLM-generated
attack-type × disguise-tier grid described in Section III-B of the paper
("Confound-Aware Evaluation of Semantic Prompt Injection Detection in Large
Language Models"). It exists so reviewers and readers can see concrete
examples of each attack type at each disguise tier without the full
malicious corpus being released (see the Ethics and Broader Impact section
of the paper for why the complete dataset is not publicly distributed).
## Contents
`illustrative_grid_sample.csv` — 88 prompts, 2 examples for each of the
44 (11 attack type × 4 disguise tier) cells in the grid, drawn from the
same generation run used to build the training/evaluation dataset.
Columns:
| Column | Meaning |
|---|---|
| `uid` | Unique identifier matching the internal dataset (not globally unique across releases) |
| `prompt` | The attack prompt text |
| `attack_type` | One of the 11 primary attack categories (Section III-A) |
| `disguise_tier` | 0–3, ordinal disguise level |
| `disguise_tier_label` | Obvious / Mild / Moderate / High |
| `injection_vector` | direct or rag_document |
| `obfuscation_method` | Surface technique used to construct the disguise |
| `topic` | Harm-domain topic label used during generation |
## Sampling method
Selected with a fixed random seed (42) from the full 1,463-row malicious
grid, 2 rows per cell. Rows whose `topic` was
`chemical, biological, or weapons synthesis` were deprioritized in favor of
another topic from the same cell where one existed, to avoid including
CBRN/weapons-adjacent content in the public sample; every cell in this
release ended up filled from a non-CBRN alternative. This is a deliberate,
bounded release — it is not a random or representative sample of the full
dataset's severity distribution, and should not be treated as such for any
purpose beyond illustrating what the attack-type × tier grid looks like.
## What is intentionally excluded
- The remaining ~1,375 rows of the malicious grid.
- The curated seed and synthetic-template malicious examples.
- The out-of-distribution probe set (`probe_data/`).
- Any downstream LLM completions or harm-judge verdicts (`results/`).
These remain available to researchers on reasonable request; see the
paper's Ethics and Broader Impact section.