flwrlabs/fed-phishing-urls

Dataset Card for Federated Phishing URLs

Dataset Summary

This dataset is a federated, non-IID phishing URL classification benchmark derived from two public Hugging Face datasets:

• ealvaradob/phishing-dataset, using the urls.json file.
• kmack/Phishing_urls, using the merged train+test+valid splits.

The resulting dataset contains URL strings, binary phishing labels, and a client_id field assigning each example to one of 100 simulated clients. Client assignment is performed before train/test splitting, so both global splits preserve per-client membership.

The dataset is intended for federated learning and non-IID robustness experiments in phishing URL detection.

Note: You may see a warning that one or more dataset files were scanned as unsafe. This is expected for a dataset containing phishing URLs. The warning reflects the presence of malicious or suspicious URL strings in the data and does not mean that the dataset files themselves are unsafe to download or use for research. Avoid opening URLs directly, and process them in a safe environment.

Dataset Structure

Data Instances

Each example has the following fields:

{
  "url": "www.example.com/login",
  "label": 0,
  "client_id": 17
}

Data Fields

Field	Type	Description
url	string	Raw URL text from the source datasets.
label	int64	Binary classification label. 1 indicates phishing or malicious; 0 indicates benign or legitimate.
client_id	int64	Simulated federated client identifier. Values range from 0 to 99 by default.

Splits

Split	Description
train	Concatenation of each client’s local training partition.
test	Concatenation of each client’s local held-out test partition.

The default generation script uses a per-client test size of 0.1, producing approximately 90% train and 10% test data for each client.

Dataset Creation

Source Data

This dataset merges URL classification samples from:

1. ealvaradob/phishing-dataset

   • The export script downloads urls.json directly from the dataset repository.
   • The source dataset describes labels as 1 for phishing and 0 for benign.

2. kmack/Phishing_urls

   • The export script loads train+test+valid from the Hugging Face dataset.
   • The source dataset uses columns text and label; text is converted into the url field in this derived dataset.

Preprocessing

The generation script performs the following preprocessing steps:

1. Loads both source datasets.
2. Merges all URL strings and labels.
3. Deduplicates examples by a normalized byte representation of each URL:
   • lowercases the URL,
   • decodes percent-encoded byte sequences,
   • compares URLs by the resulting byte string.
4. Extracts a coarse feature bucket for each URL, such as:
   • ip_host,
   • shortener,
   • suspicious_tld,
   • query_heavy,
   • deep_subdomain,
   • very_long,
   • digit_heavy,
   • symbol_heavy,
   • randomish_host,
   • standard.
5. Allocates examples to clients using a non-IID client-first procedure.
6. Splits each client’s local data into train and test partitions.
7. Concatenates all local train partitions into the global train split and all local test partitions into the global test split.

Federated Partitioning

The default configuration creates 100 simulated clients with both quantity skew and distribution skew.

Quantity Skew

Client sizes are sampled from a log-normal distribution with a minimum client size. The default parameters are:

Parameter	Default
num_clients	100
min_client_size	50
size_sigma	1.0

Label Skew

Each client receives a different phishing prevalence tendency sampled from a beta distribution. The default label skew parameter is:

Parameter	Default
label_alpha	0.5

Lower values of label_alpha produce stronger label imbalance across clients.

Feature Skew

Each client is assigned preferences over URL feature buckets. During allocation, examples are assigned based on client capacity, label preference, and feature preference. The default feature-skew parameters are:

Parameter	Default
feature_temperature	0.35
preferred_features_per_client	3

This makes some clients more likely to contain certain URL patterns, such as shorteners, suspicious TLDs, deep subdomains, or query-heavy URLs.

Intended Uses

This dataset is suitable for:

• federated binary classification,
• cross-client generalization studies,
• non-IID robustness benchmarking,
• phishing URL detection,
• cybersecurity model evaluation,
• client-level distribution shift analysis.

Out-of-Scope Uses

This dataset should not be used to:

• identify or accuse real users or organizations of phishing activity,
• make high-stakes security decisions without additional validation,
• crawl, visit, or interact with URLs directly,
• deploy a production phishing detector without current threat-intelligence evaluation.

URLs may be stale, inactive, benignly repurposed, or no longer representative of current phishing behavior.

Data Splitting

The dataset uses a client-first splitting strategy:

1. Assign each deduplicated URL to a simulated client.
2. Split each client’s local data into train/test.
3. Concatenate local partitions into global splits.

This avoids leakage where a client appears only in one split and ensures every split keeps client identifiers available for federated evaluation.

When possible, each client’s local train/test split is stratified by the binary label. If a client has too few samples from one class, the script falls back to a random split for that client.

Loading the Dataset

from datasets import load_dataset

ds = load_dataset("flwrlabs/fed-phishing-urls")
print(ds)
print(ds["train"][0])

A simple per-client subset can be selected with:

from flwr_datasets import FederatedDataset
from flwr_datasets.partitioner import NaturalIdPartitioner

fds = FederatedDataset(
    dataset="flwrlabs/fed-phishing-urls",
    partitioners={"train": NaturalIdPartitioner(partition_by="client_id")},
)
partition = fds.load_partition(partition_id=0)

Label Semantics

Label	Meaning
0	Benign / legitimate URL
1	Phishing / malicious URL

The label semantics follow the upstream phishing URL datasets.

Limitations

• The dataset is derived from public URL datasets and may include outdated or inactive URLs.
• Labels reflect the upstream datasets and may contain noise.
• Feature buckets are heuristic and intended only for partitioning; they are not included as dataset columns.
• URL distributions and phishing tactics change over time, so results on this dataset may not reflect current real-world detection performance.

Ethical Considerations

This dataset contains URLs associated with phishing and malicious activity. Users should avoid visiting URLs directly and should process the data in a safe, sandboxed environment. The dataset is intended for research and benchmarking, not for direct operational threat blocking without further validation.

Citation

Please cite the upstream datasets when using this derived dataset. If you're using this dataset with Flower Datasets, you can cite Flower.

@misc{alvarado_phishing_dataset,
  author = {Alvarado, E.},
  title  = {{Phishing Dataset}},
  year   = {n.d.},
  url    = {https://huggingface.co/datasets/ealvaradob/phishing-dataset}
}

@misc{kmack_phishing_urls,
  author = {kmack},
  title  = {{Phishing URLs Dataset}},
  year   = {2024},
  url    = {https://huggingface.co/datasets/kmack/Phishing_urls}
}

@article{DBLP:journals/corr/abs-2007-14390,
  author       = {Daniel J. Beutel and
                  Taner Topal and
                  Akhil Mathur and
                  Xinchi Qiu and
                  Titouan Parcollet and
                  Nicholas D. Lane},
  title        = {Flower: {A} Friendly Federated Learning Research Framework},
  journal      = {CoRR},
  volume       = {abs/2007.14390},
  year         = {2020},
  url          = {https://arxiv.org/abs/2007.14390},
  eprinttype    = {arXiv},
  eprint       = {2007.14390},
  timestamp    = {Mon, 03 Aug 2020 14:32:13 +0200},
  biburl       = {https://dblp.org/rec/journals/corr/abs-2007-14390.bib},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
}

Dataset Card Contact

In case of any doubts, please contact Flower Labs.