agarw347/VERA-Dataset

VERA EMBER2024-compatible Features

EMBER2024-compatible static features for every binary in VERA (Validated Execution-based Ransomware Analysis), a dual-validated, behaviorally-verified dataset of 15,000+ raw Windows ransomware binaries. No binaries are included; this repository contains only features and labels.

Because the format is identical to EMBER2024's, VERA samples can be combined with EMBER2024 data, or EMBER2024-trained models can be tested on VERA ransomware.

Links

• GitHub: https://github.com/garvit-agarwal-purdue/VERA-Dataset
• Zenodo (full dataset): https://doi.org/10.5281/zenodo.17904705
• Paper: https://doi.org/10.1109/ICCCN69946.2026.11662597
• EMBER2024: https://github.com/FutureComputing4AI/EMBER2024

How the features were made

Features were generated with EMBER2024's thrember at commit 0ef753e81d98bf209f71b03cd331dfc190b5b54d, EMBER feature version 3, giving 2,568-dimensional vectors. The binaries are only read for static parsing; nothing is executed. The generation script is in EMBER_Features/ of the GitHub repository. See run_info.json for run details.

Contents

Partition	Samples	Non-PE
Active	15,448	312
Latent	22,837	6

Total: 38,285 samples, 0 failures. Non-PE files (scripts, MSI/OLE, LNK, ELF, DEX and blobs) are kept and flagged with is_pe = false. For these, only the general, byte-histogram, byte-entropy and string features are populated; the PE-specific blocks are empty or zero.

Columns

data/ember_vectors.parquet (config vectors, split train), one row per sample:

• sha256 (string): SHA256 of the VERA binary.
• partition (string): Active or Latent. Active: samples that satisfy static consensus and exhibit at least one high-confidence ransomware behavior during execution, such as volume shadow copy deletion, high-entropy file writes, or ransom note generation. Latent: samples that satisfy the same static consensus but stay behaviorally silent within the execution window, for reasons such as defunct infrastructure, delayed triggers, or environment-specific execution checks.
• family (string, null when empty): final unified family label. Only available for Active samples.
• label_source (string, null when empty): consensus is the Stage I tool consensus; model_predicted is a Stage II XGBoost propagation (Contested Set). Only available for Active samples.
• is_pe (bool): whether the file was parsed as a PE.
• features (list of float32, length 2,568): EMBER2024 feature vector.

Active/Latent are not EMBER malicious/benign labels; define your own target from partition.

Usage

import numpy as np
from datasets import load_dataset

ds = load_dataset("agarw347/VERA-Dataset", split="train")

X = np.asarray(ds["features"], dtype=np.float32)   # (38285, 2568)
y = np.array([p == "Active" for p in ds["partition"]])

Raw EMBER-style records (raw/raw_features.jsonl.gz, one JSON object per sample, with vera_sha256, vera_partition, vera_family and vera_label_source added) are a plain file download for EMBER tooling. They are not a datasets config because the nested JSON is heterogeneous:

import gzip, json
from huggingface_hub import hf_hub_download

p = hf_hub_download("agarw347/VERA-Dataset", "raw/raw_features.jsonl.gz", repo_type="dataset")
with gzip.open(p, "rt") as f:
    records = [json.loads(line) for line in f]

To combine with EMBER2024, stack X with EMBER2024's vectorized features (same 2,568-dimensional format) and define your own labels.

License

CC BY 4.0 for the features and labels. The binaries are not distributed here.

Citation

If you use VERA in your work, please cite both the paper and the dataset.

Paper:

@INPROCEEDINGS{11662597,
  author={Agarwal, Garvit and Vasanth, Nithish and Cho, Seunghyun and Alomayri, Yousef Mohammed Y. and Pumphrey, Noah D. and Li, Feng and Xie, Yucheng and Zou, Xukai},
  booktitle={2026 35th International Conference on Computer Communications and Networks (ICCCN)}, 
  title={VERA: A Dual-Validated, Behaviorally-Verified Repository of 15,000+ Raw Windows Ransomware Binaries}, 
  year={2026},
  volume={},
  number={},
  pages={1-9},
  keywords={Labeling;Ransomware;Modeling;Malware;Signal detection;Printing;Tools;Windows;Machine learning;Pipelines;Ransomware;Dataset;Machine Learning Security;Dynamic Analysis;Reproducibility},
  doi={10.1109/ICCCN69946.2026.11662597}
}

Dataset:

@dataset{agarwal2026vera_dataset,
  title     = {VERA: Validated Execution-based Ransomware Analysis Dataset},
  author    = {Agarwal, Garvit and Vasanth, Nithish and Cho, Seunghyun and Alomayri, Yousef Mohammed Y. and Pumphrey, Noah D. and Li, Feng and Xie, Yucheng and Zou, Xukai},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.17904705},
  url       = {https://doi.org/10.5281/zenodo.17904705}
}