LAMDA: A Longitudinal Android Malware Dataset for Drift Analysis

This dataset contains a longitudinal benchmark for Android malware detection designed to analyze and evaluate concept drift in machine learning models. It includes labeled and feature-engineered Android APK data from 2013 to 2025 (excluding 2015), with over 1 million samples collected from real-world sources.

Dataset Details

LAMDA is the largest and most temporally diverse Android malware dataset to date. It enables studies in concept drift, temporal generalization, family classification, and robust malware detection over time. Each sample includes static features (from .data files), metadata (VirusTotal detections, family name, timestamp), and binary labels.

The dataset was created using ~1M APKs from AndroZoo, with additional metadata and labels derived via VirusTotal and AVClass2. Labels are assigned using a 4+ AV detection threshold to reduce noise.

Dataset Sources

Repository: https://huggingface.co/datasets/IQSeC-Lab/LAMDA
Project Website: https://iqsec-lab.github.io/LAMDA/
Paper: https://arxiv.org/abs/2505.18551

Uses

Direct Use

Malware classification
Family prediction
Concept drift analysis
Temporal generalization benchmarks
SHAP-based feature attribution drift analysis
Continual learning evaluation (e.g., class-IL, replay)

Out-of-Scope Use

Dynamic behavior analysis (no runtime traces)
On-device malware detection (model integration not provided)

Dataset Structure

Each year is stored in a subdirectory:

2013/
├── 2013_train.parquet
├── 2013_test.parquet
...
2025/
├── 2025_train.parquet
├── 2025_test.parquet

Each .parquet contains:

Column	Description
`label`	0 = benign, 1 = malware
`family`	Malware family name (via AVClass2)
`vt_count`	VirusTotal vendor detection count
`year_month`	Timestamp in YYYY-MM format
`feat_0 ... feat_4560`	Static bag-of-words features (int8)
`hash`	Sample SHA256 hash (used as index)

A feature_mapping.csv maps each feat_i to its original static token.

Dataset Creation

Curation Rationale

To enable longitudinal and realistic evaluation of ML-based malware detection systems that must remain effective in the face of temporal and adversarial drift.

Source Data

APK samples were downloaded from AndroZoo and processed using static analysis to extract .data files. Metadata was merged from a curated CSV containing VirusTotal counts and family assignments via AVClass2.

Data Collection and Processing

Extracted feature vectors from .data files (comma-separated tokens)
Labeled malware if vt_detection ≥ 4
Assigned families via AVClass2
Feature vectors vectorized using bag-of-words (sparse)
Feature selection via VarianceThreshold=0.001 → 4,561 features
Train/test split (80/20) stratified by label, year-wise

Who are the source data producers?

Original APKs are from AndroZoo. Annotations and processing were conducted by IQSeC Lab at the University of Texas at El Paso.

Annotations

Annotation Process

Malware/benign labels based on AV vendor threshold (≥4)
Family labels from AVClass2
All annotations generated using automated pipelines

Who are the annotators?

Researchers at IQSeC Lab via static tooling and AVClass2

Personal and Sensitive Information

No PII or private user data is included. APKs are anonymized binaries.

Bias, Risks, and Limitations

Biased toward highly detected malware (AV-centric labeling)
No dynamic/runtime behavior
Concept drift is dataset-driven, not simulation-based

Recommendations

Normalize class balance before training
Use continual or time-aware validation schemes
SHAP explanations should be anchored year-wise

Citation

BibTeX:

@article{lamda,
  title     = {{LAMDA: A Longitudinal Android Malware Benchmark for Concept Drift Analysis}},
  author    = {Md Ahsanul Haque and Ismail Hossain and Md Mahmuduzzaman Kamol and Md Jahangir Alam and Suresh Kumar Amalapuram and Sajedul Talukder and Mohammad Saidur Rahman},
  year      = {2025},
  eprint    = {2505.18551},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CR},
  url       = {https://arxiv.org/abs/2505.18551}
}