ahmedBargady/MIAFF_Dataset

MIAFF Dataset (Modular Infrastructure-Aware Fusion Framework)

Standardized domain detection benchmarks paired with 135 extracted infrastructure metadata features.

1. Tasks & Datasets

Each dataset benchmark is partitioned into train.csv (60%), dev.csv (20%), and test.csv (20%), stratified by class/label:

• Malware_Binary/ (Binary malware domain classification: legit vs. malicious)
  • deepurlbench/
  • domainradar/
  • macformer/
  • threatfox/
• Malware_Family/ (Multi-class malware family attribution)
  • domainradar/ (61 classes)
  • threatfox/ (65 classes)
• Phishing_Binary/ (Binary phishing domain classification: legit vs. malicious)
  • deepurlbench/
  • domainradar/
  • macformer/

2. Dataset Columns (139 Total)

Identification & Target Columns (4)

Columns excluding the tabular features, ordered at indices 0–3:

Column	Type	Description
input	string	Domain name evaluated
ip	string	Resolved IP address / A-records
class	string	Benchmark-specific class label (e.g., specific malware family or sub-class)
label	string	Normalized evaluation label (legit or malicious)

Feature Columns (135)

The remaining 135 columns represent infrastructure metadata signals grouped across 5 categories:

• RDAP (rdap_*, 37 features): Registration, registrar, domain age, expiration, update recency.
• DNS (dns_*, 29 features): Nameserver distribution, record counts, TTL statistics.
• TLS (tls_*, 26 features): Certificate validity periods, cipher suites, issuer properties.
• IP (ip_*, 31 features): ASN attributes, geolocation, CIDR routing characteristics.
• Cross-signals (x_*, 12 features): Ratios and interactions (e.g., TTL-to-age ratios, new domain flags).

3. Features Reference (features_description.csv)

Every feature is indexed and documented in features_description.csv with 3 schema fields:

• feature_name: The verbatim column name in the CSV files.
• category: Feature subsystem (RDAP, DNS, TLS, IP, or Cross-signals).
• description: Plain-text technical definition of the extracted feature.

4. Metadata Feature Selection

Candidate features and data isolation

The DomainRadar EDC tool produced 137 numeric candidate signals per domain. The published MIAFF tables retain the 135 infrastructure metadata features documented in features_description.csv; the EDC input-validation indicators input_valid and input_is_ip_address are not part of this published metadata schema. Identifiers and targets were not considered metadata features.

Feature selection was performed independently for every source dataset because missingness, value distributions, and predictive utility can differ across data sources. Only train.csv and dev.csv were used: the training split determined feature quality, relevance, and stability, while the development split was used for the final binary-task safeguard. The test split was never used to select or finalize features.

Per-dataset selection protocol

For each dataset, the selector applied the following fixed protocol:

1. Statistical quality filtering. A feature was removed when it was constant, had at least 95% missing values, or one non-missing value accounted for at least 99.5% of its observations. Remaining missing values were median-imputed.
2. Predictive-relevance filtering. Mutual information with the task target was estimated on the training data. Features with mutual information of at least 0.001 were retained; variables with at most 20 distinct values were treated as discrete during estimation.
3. Stability selection. Quality and relevance filtering were repeated inside five stratified training folds using a fixed random seed of 42. A feature was stable when it was selected in at least four of the five folds (selection frequency $\geq 0.80$).
4. Binary-task safeguard. For malware and phishing binary classification, the stable subset was compared with the complete candidate set using median imputation, standardization, and class-balanced logistic regression. The stable subset was frozen only when its development macro-F1 was no more than 0.005 below that of the complete set; otherwise, all candidates were retained. Balanced accuracy, MCC, AUROC, and average precision were recorded for auditing but did not control the decision.
5. Multiclass finalization. For malware-family attribution, the recurrent stable subset was frozen directly. The multiclass path did not use the logistic-regression development safeguard.

Every run records the selected feature list, per-feature quality and stability audit, cross-validation summary, thresholds, final decision, and confirmation that test data were not used for finalization.

Task-level feature sets

Dataset-specific selections were combined by union, not intersection. For task $t$ with datasets $d$, the published task set is

$$F_t=\left(\bigcup _d{F}_{t,d}\right)\cap F_{\mathrm{p}\mathrm{u}\mathrm{b}\mathrm{l}\mathrm{i}\mathrm{s}\mathrm{h}\mathrm{e}\mathrm{d}},F\_t\; =\; \backslash left(\backslash bigcup\_d\; F\_\{t,d\}\backslash right)\; \backslash cap\; F\_\{\backslash mathrm\{published\}\},$$

where $F_{\mathrm{published}}$ is the 135-feature metadata schema. The union preserves a feature when it is selected for at least one dataset in the task and gives all datasets for that task a common input schema.

Task	Final per-dataset selections	Published task union	RDAP / DNS / TLS / IP / Cross
Malware binary	ThreatFox: 88; MACFormer: 86; DeepURLBench: 78; DataInBriefHranicky: 76	108	27 / 27 / 24 / 21 / 9
Phishing binary	MACFormer: 137; DeepURLBench: 137; DataInBriefHranicky: 137	135	37 / 29 / 26 / 31 / 12
Malware family	ThreatFox: 104; DataInBriefHranicky: 102	115	31 / 26 / 24 / 24 / 10

For phishing, the binary safeguard retained all 137 EDC candidates in each individual dataset; intersecting their union with the published metadata schema removed the two input-validation indicators and produced the 135-feature task set. The packaged task lists are Malware_Binary/malware.txt, Phishing_Binary/phishing.txt, and Malware_Family/malware_family.txt. Each file defines the common metadata input schema for its task; training loads only the columns named in that file rather than all metadata columns indiscriminately.

5. License & Research Use

This dataset is licensed under the MIT License, permitting open use, modification, and distribution for academic research and evaluation.