Update README.md

README.md

@@ -1,3 +1,51 @@
----
-license: apache-2.0
----
+---
+license: apache-2.0
+---
+
+# EMBER2024 Dataset
+
+EMBER2024 is an update to the [EMBER2017 and EMBER2018](https://github.com/elastic/ember/) datasets. It includes raw features and labels for 3.2 million malicious and benign files from 6 different file types (Win32, Win64, .NET, APK, ELF, and PDF). EMBER2024 is meant to allow researchers to explore a variety of common malware analysis classification tasks. The dataset includes 7 types of labels and tags that support malicious/benign detection, malware family classification, malware behavior prediction, and more.
+
+Paper about the classifiers: [https://arxiv.org/pdf/2506.05074](https://arxiv.org/pdf/2506.05074)
+
+Instructions for downloading EMBER2024 and training classifiers: [https://github.com/FutureComputing4AI/EMBER2024/](https://github.com/FutureComputing4AI/EMBER2024/)
+
+
+## EMBER2024 Contents
+
+EMBER2024 includes features and labels for malware that was first uploaded to VirusTotal between Sep. 24th, 2023 and Dec. 14th, 2024. There are exactly 50,500 files chosen from each week of that time period, with the first 52 weeks of files making up the training set and the last 12 going to the test set. This lets researchers simulate how effectively a classifier might detect malware that is newer than its training corpus. In total, the training set is 2,626,000 files and the test set is 606,000 files.
+
+
+#### File Statistics
+
+| File Type   | Malicious + Benign (Weekly) | Train Total | Test Total |
+| -------- | ------- | ------ | ------- |
+| Win32  | 30,000    | 1,560,000 | 360,000 |
+| Win64 | 10,000     | 520,000 | 120,000 |
+| .NET    | 5,000    | 260,000 | 60,000 |
+| APK  | 4,000    | 208,000 | 48,000 |
+| PDF | 1,000  | 52,000 | 12,000 |
+| ELF    | 500    | 26,000 | 6,000 |
+
+
+#### Challenge Set
+
+EMBER also includes features and labels for 6,315 malicious files in a "challenge set". These files initially went undetected by ~70 antivirus products on VirusTotal but were later found to be malicious. The challenge set is an excellent resource for assessing how well a machine larning classifier is able to detect evasive malware.
+
+
+## EMBER Feature Version 3
+
+The previous EMBER feature versions were pinned to [LIEF](lief.re) version 0.9.0, which requires Python 3.6. EMBER feature version 3 ("thrember") is a re-implementation of the EMBER feature vector format that uses the [pefile](https://github.com/erocarrera/pefile) library instead. pefile is stable and has no dependencies, making it ideal going forward. We have also made several addition to the EMBER feature vector format, which now includes features from the DOS header, Rich header, PE data directories, Authenticode signatures, and warnings during PE parsing. Furthermore, we have added support for feature extraction from non-PE files using a subset of the EMBER feature version 3 format. We show that effective classifiers for APK, ELF, and PDF files can be trained using just features from general file info, byte statistics, and string statistics.
+
+
+## Citing
+
+If you use EMBER2024 in your own research, please cite it using:
+
+```
+@inproceedings{joyce2025ember,
+      title={EMBER2024 - A Benchmark Dataset for Holistic Evaluation of Malware Classifiers},
+      author={Robert J. Joyce and Gideon Miller and Phil Roth and Richard Zak and Elliott Zaresky-Williams and Hyrum Anderson and Edward Raff and James Holt},
+      year={2025},
+      booktitle={Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
+}