File size: 4,875 Bytes
cf181a5 fb2219f cf181a5 fb2219f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 | ---
license: gpl-3.0
base_model: FacebookAI/roberta-base
datasets:
- CIRCL/vulnerability-attack-techniques
language:
- en
tags:
- security
- vulnerability
- mitre-attack
- cve
- bi-encoder
- sentence-similarity
library_name: transformers
---
# vulnerability-attack-technique-biencoder
A label-semantics bi-encoder that suggests MITRE ATT&CK (Enterprise)
techniques for a CVE by scoring the vulnerability description against the
**official ATT&CK technique descriptions** in a shared embedding space.
Unlike the companion classification head
([`CIRCL/vulnerability-attack-technique-classification-roberta-base`](https://huggingface.co/CIRCL/vulnerability-attack-technique-classification-roberta-base)),
it can rank *any* technique that has an official description — the label
is text, not a learned output row.
One shared `roberta-base` encoder embeds both the CVE text
(title + description) and each technique's STIX name+description
(citation markup stripped, 256 tokens), mean-pooled and L2-normalized;
the score is a learned affine over the cosine. Trained on the curated
gold set [`CIRCL/vulnerability-attack-techniques`](https://huggingface.co/datasets/CIRCL/vulnerability-attack-techniques)
(~1,200 CVEs, CTID methodology) with per-label-weighted BCE over a
53-parent-technique vocabulary, with VulnTrain
(`vulntrain-train-attack-biencoder`).
## When to use which model
- **Classification head**: best top-5 ranking on the trained vocabulary
(recall@5 0.667 ± 0.015 across five seeds).
- **This bi-encoder**: slightly lower recall@5 (0.643 ± 0.019) but the
largest consistent rare-technique gain measured on this task
(macro-F1 0.212 ± 0.011 vs 0.176 ± 0.016, +21% relative), and
open-vocabulary ranking over all 222 active parent techniques
(recall@5 0.515 ± 0.020, 2.3× a generic zero-shot sentence embedder).
Caveat measured in the accompanying paper: zero-shot ranking of
techniques *absent from training* does **not** benefit from this
fine-tuning — in a five-fold label-holdout evaluation the fine-tuned
encoder ranked held-out techniques below a generic MiniLM embedder.
Rankings for techniques outside the 53-technique training vocabulary
should be treated as no better than generic semantic similarity.
## Usage
The repository ships `technique_texts.json` (the exact technique texts
used at training time) and the scoring calibration in
`config.biencoder`:
```python
import json, torch
from huggingface_hub import hf_hub_download
from transformers import AutoModel, AutoTokenizer
model_id = "CIRCL/vulnerability-attack-technique-biencoder"
tokenizer = AutoTokenizer.from_pretrained(model_id)
encoder = AutoModel.from_pretrained(model_id).eval()
cfg = encoder.config.biencoder
texts = json.load(open(hf_hub_download(model_id, "technique_texts.json")))
def embed(batch, max_length=512):
enc = tokenizer(batch, padding=True, truncation=True,
max_length=max_length, return_tensors="pt")
hidden = encoder(**enc).last_hidden_state
mask = enc["attention_mask"].unsqueeze(-1)
pooled = (hidden * mask).sum(1) / mask.sum(1)
return torch.nn.functional.normalize(pooled, dim=-1)
techniques = sorted(texts)
with torch.no_grad():
technique_emb = embed([texts[t] for t in techniques],
cfg["technique_max_length"])
cve_emb = embed(["Improper neutralization of special elements used "
"in an OS command in the web management interface..."])
scores = cfg["logit_scale"] * (cve_emb @ technique_emb.T) + cfg["logit_bias"]
for idx in scores[0].topk(5).indices:
print(techniques[idx], float(scores[0][idx]))
```
Evaluation and stratified breakdowns are reproducible with
`vulntrain-validate-attack-classification --method biencoder --model
CIRCL/vulnerability-attack-technique-biencoder` (add `--candidates full`
for open-vocabulary ranking over all active parent techniques).
## Intended use and limitations
The model generates **candidate techniques for analyst review**, not
authoritative mappings. Technique-to-CVE mapping involves analyst
judgment; the training labels inherit the CTID methodology's
subjectivity, and the gold set over-represents exploited and enriched
CVEs. English descriptions only; parent-level techniques only.
## References
- Bonhomme, C., & Dulaunoy, A. (2026). *Mapping CVEs to MITRE ATT&CK
Techniques: A Curated Gold-Set Classifier and the Limits of
LLM-Assisted Label Expansion.* [arXiv:2607.25572](https://arxiv.org/abs/2607.25572)
- Bonhomme, C., & Dulaunoy, A. (2026). *Beyond the Description:
Structured Metadata and Label Semantics for CVE-to-ATT&CK Mapping.*
(follow-up paper, in preparation — source of all numbers above)
- Trained with [VulnTrain](https://github.com/vulnerability-lookup/VulnTrain)
as part of the [Vulnerability-Lookup](https://vulnerability.circl.lu) project.
|