Appendix: MELD-DS-448 Dataset Overview

Dataset Overview

MELD-DS-448 contains 26,166 malicious samples spanning 448 distinct malware families collected from April 2020 to August 2025. All samples are uniquely identified by SHA-256 hashes and include precise "First Seen" timestamps.

Family Distribution Characteristics: The dataset exhibits a typical long-tail distribution, with 35.7% singleton families (only 1 sample) and 64.7% small-scale families (≤5 samples). Head concentration is significant, with the top 5 families covering 30.1% of samples and the top 10 families covering 41.5% of samples. Major families include LummaStealer (2,966 samples, 11.3%), Formbook (2,091 samples, 8.0%), SnakeKeylogger (1,045 samples, 4.0%), and others.

Temporal Evolution Patterns: Samples are primarily concentrated in 2024-2025 (96.8%), with the majority appearing in 2025 (77.1% of samples). The temporal distribution shows rapid growth in recent years, with 2024 contributing 19.7% and earlier years contributing minimal samples. This demonstrates the rapid evolution characteristics of contemporary malware ecosystems.

Dataset Statistical Overview

Metric	Value	Description
Total Families	448	Total number of distinct malware families
Total Samples	26,166	Total number of malicious samples
Avg Samples/Family	58.4	Average samples per family
Sample Count Median	3	Median of family sample counts
Singleton Families	160 (35.7%)	Families with only 1 sample
Small Families	290 (64.7%)	Families with ≤5 samples
Large Families	36 (8.0%)	Families with ≥100 samples
Top 5 Coverage	30.1%	Sample coverage by top 5 families
Top 10 Coverage	41.5%	Sample coverage by top 10 families

Annual Evolution Statistics

Year	Active Families	Samples	Percentage
2020	-	6	0.0%
2021	-	258	1.0%
2022	-	343	1.3%
2023	-	227	0.9%
2024	-	5,146	19.7%
2025	-	20,186	77.1%

Standardized Analysis Artifacts

Each sample in MELD-DS-448 provides four types of standardized analysis data generated through unified CAPE Sandbox analysis in virtualized Windows 10 x64 (22H2) environments:

1. CAPE JSON Reports - Complete structured analysis results containing behavioral indicators, network activities, file system operations, registry modifications, and process execution traces, as the original analysis reports from CAPEv2.

2. Markdown Structured Reports - Converting CAPE JSON reports into LLM-friendly structured Markdown format containing complete behavioral events, API call patterns, process tree information, and temporal analysis. These reports are specifically designed for large language model processing and understanding.

3. API Call Sequences - Chronologically ordered sequences of Windows API function calls captured during dynamic execution, including parameters and return values, converted from CAPEv2's JSON reports. These sequences enable fine-grained behavioral modeling and sequence-based machine learning approaches.

4. ASM Disassembly Files - Static disassembly output providing low-level instruction sequences and control flow information. These artifacts support static analysis techniques and hybrid approaches combining static and dynamic features.

Note on ASM File Coverage: Out of 26,166 total samples, 361 samples (1.38%) do not have corresponding ASM disassembly files due to disassembly process failures during reverse engineering analysis. These missing files are documented in asm_loss.csv for reference. The remaining 25,805 samples (98.62%) have complete ASM disassembly data available.

Data Quality and Coverage

All 26,166 samples (100% coverage) include complete metadata and three primary analysis artifact types (CAPE JSON reports, Markdown reports, and API call sequences). ASM disassembly files are available for 25,805 samples (98.62%), with 361 samples missing ASM files due to disassembly process failures. The dataset ensures sample uniqueness through SHA-256 deduplication and maintains temporal consistency with verified timestamps. File sizes range from 87.3 KB to 301.3 MB (median: 3.6 MB), with the complete dataset totaling 479 GB of analysis artifacts and metadata.

Dataset File Structure

The dataset files are organized in the Dataset/ directory with large files split into volumes for easier download and Git LFS compatibility:

File Restoration Instructions

Due to file size limitations, large dataset files have been split into 4GB volumes. To restore the original files, use the following commands:

1. ASM Disassembly Files (27GB total)

7z x asm.7z.001

2. API Call Sequences (8.9GB total)

7z x api_sequence.7z.001

3. CAPE JSON Reports (8.5GB total)

7z x cape_reports.7z.001

4. Markdown Reports (67MB - no splitting needed)

• File: cape_reports_malicious_md.7z
• Can be extracted directly: 7z x cape_reports_malicious_md.7z

Requirements

• 7-Zip: Required for extracting split archives
• Disk Space: Ensure at least 500GB free space for extraction
• Memory: Recommended 8GB+ RAM for processing large files

License

This project is licensed under the CC BY-NC-SA 4.0 (Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International) license.