Olaroti/phishing-dataset

Phishing Detection Dataset

A balanced, feature-engineered dataset for phishing detection research containing 10,000 labelled indicators (5,000 phishing / 5,000 benign) with 38 hand-crafted lexical features derived from phishing detection literature.

Dataset Summary

Property	Value
Total Examples	10,000
Phishing (label=1)	5,000
Benign (label=0)	5,000
Features	38 engineered + 5 base columns
Indicator Types	URL, domain, hostname, IPv4
Format	CSV
Size	~1.3 MB

Data Sources

Source	Description	Rows
AlienVault OTX	Open Threat Exchange pulses matching phishing-related queries (phishing, phish, credential theft, smishing)	8,000 raw
Tranco List	Research-grade top domains (aggregated Alexa / Umbrella / Majestic / Cloudflare Radar)	6,000 raw

Feature Categories

Base Columns

Column	Type	Description
indicator	string	Raw indicator value
indicator_type	string	URL / domain / hostname / IPv4
source	string	OTX or tranco
pulse_id	string	OTX pulse ID (phishing rows only)
label	int	1 = phishing, 0 = benign

Lexical Features (20)

URL structure and character-level features derived from phishing detection literature:

Feature	Description
url_length	Total character length
hostname_length	Hostname component length
path_length	Path length
query_length	Query string length
num_dots, num_hyphens, num_at, num_question, num_equals, num_underscore, num_ampersand, num_tilde, num_percent, num_slash	Character counts
num_digits, digit_letter_ratio	Digit analysis
num_subdomains	Subdomain label count
longest_token_length, avg_token_length	Token statistics
hostname_entropy	Shannon entropy of hostname

Host-Based Features (9)

Feature	Description
has_https	HTTPS scheme present
has_port	Non-default port
has_ip_in_hostname	IP literal in hostname
has_punycode	IDN homograph (xn--) present
has_at_symbol	@ present (auth confusion)
has_double_slash_in_path	Path contains //
tld, tld_length	Effective TLD
is_suspicious_tld	TLD in phishing-prone list

Content / Brand Features (3)

Feature	Description
num_suspicious_keywords	Count of {login, secure, account, verify, banking, password, ...}
brand_in_subdomain	Brand token in subdomain
brand_in_path	Brand token in path

Label Strategy

• Phishing (1): OTX pulses matching queries phishing, phish, credential theft, smishing
• Benign (0): Top-ranked Tranco domains
• Leakage Prevention: Benign rows whose registrable domain (eTLD+1) appears in the phishing set are dropped before sampling
• Balance: 50/50 split (default), configurable via --benign-ratio

Usage

With Hugging Face datasets

from datasets import load_dataset

dataset = load_dataset("999Roti/phishing-dataset")
df = dataset["train"].to_pandas()

# Access features and labels
X = df.drop(columns=["indicator", "indicator_type", "source", "pulse_id", "label"])
y = df["label"]

With pandas

import pandas as pd

df = pd.read_csv("phishing_dataset.csv")
print(df["label"].value_counts())  # 5000 / 5000

Baseline Experiment

A quick Random Forest baseline on a train/test split:

from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import train_test_split
from sklearn.metrics import classification_report
import pandas as pd

df = pd.read_csv("phishing_dataset.csv")

# Select numeric feature columns
feature_cols = [c for c in df.columns if c not in
                ["indicator", "indicator_type", "source", "pulse_id", "label", "tld"]]
X = df[feature_cols]
y = df["label"]

X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size=0.2, random_state=42, stratify=y
)

clf = RandomForestClassifier(n_estimators=200, random_state=42, n_jobs=-1)
clf.fit(X_train, y_train)

print(classification_report(y_test, clf.predict(X_test)))

Expected performance: ~95-97% accuracy, ~95% F1-score.

Data Quality & Limitations

Consideration	Detail
Reproducibility	OTX data varies over time as new pulses are published; Tranco list updates daily. Seed (42) fixes sampling.
Offline features only	No network-dependent features (WHOIS, DNS, TLS, page content) — deterministic & privacy-preserving.
Indicator types mixed	URLs, domains, hostnames, IPv4 are combined; consider stratifying by indicator_type for fair evaluation.
English-centric	Brand tokens and suspicious keywords are English-focused.
Historical data	OTX pulses reflect reported threats; zero-day phishing may not be represented.

Citations

If you use this dataset, please cite the underlying feature engineering literature:

@misc{roti2026phishingdataset,
  title = {Phishing Detection Dataset},
  author = {Roti, Olaoluwa},
  year = {2026},
  howpublished = {\url{https://huggingface.co/datasets/999Roti/phishing-dataset}}
}

Feature schema derived from:

• Mohammad, R. M., Thabtah, F., & McCluskey, L. (2014). Predicting phishing websites based on self-structuring neural network. Neural Computing and Applications, 25(2), 443–458.
• Sahingoz, O. K., Buber, E., Demir, O., & Diri, B. (2019). Machine learning based phishing detection from URLs. Expert Systems with Applications, 117, 345–357.
• Marchal, S., Saari, K., Singh, N., & Asokan, N. (2016). Know Your Phish: Novel Techniques for Detecting Phishing Sites and their Targets. ICDCS.

License

MIT License. See repository for full text.

Repository

Full pipeline code: github.com/999Roti/phishing-dataset

Includes:

• src/otx_ingest.py — OTX phishing indicator ingestion
• src/benign_source.py — Tranco benign domain collection
• src/features.py — Feature extraction engine
• src/build_dataset.py — Dataset construction orchestrator