DB-Forensic-Hybrid Dataset

A large-scale, event-level MySQL database forensics benchmark dataset for multi-class attack detection and attribution using real MySQL log streams. Events are labeled at the individual log-event level against 14 MITRE ATT&CK techniques plus a Normal class, collected across 8 realistic attack scenarios.

This dataset accompanies the paper:

"DB-Forensic-Hybrid: A Hybrid Deep Learning Framework for Database Intrusion Detection and Forensic Attribution" Hyeonseop Shin — submitted to Forensic Science International: Digital Investigation (FSI:DI)

Dataset Summary

Property	Value
Task	Multi-class event-level sequence labeling
Classes	15 (Normal + 14 MITRE ATT&CK techniques)
Total events	~14.5M log events
Log sources	MySQL general log, error log, slow query log, binary log
Attack scenarios	8 (S01–S08), Leave-One-Scenario-Out splits provided
Feature dimensionality	46-dim float32 vectors (pre-extracted)
License	CC BY 4.0

---

Access

This dataset is gated — please request access via the button above. Access is manually approved. Approved users can download with:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="golfoscar/db-forensic-hybrid",
    repo_type="dataset",
    local_dir="data/",
    token="YOUR_HF_TOKEN",
)

To download only the pre-extracted features (~2.2 GB, sufficient to re-run all evaluations without GPU):

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="golfoscar/db-forensic-hybrid",
    repo_type="dataset",
    local_dir="data/",
    allow_patterns=["features_v3/*", "val_versions/*"],
    token="YOUR_HF_TOKEN",
)

---

Repository Structure

golfoscar/db-forensic-hybrid/
├── labeled/                        Raw labeled log events (Parquet)
│   ├── train/events.parquet        10.0 GB — 14,497,692 events
│   ├── val/events.parquet           1.7 GB —  2,250,771 events
│   └── test/events.parquet          1.7 GB —  1,919,768 events (held-out)
│
├── features_v3/                    Pre-extracted 46-dim features (NumPy)
│   ├── train_features.npy           1.2 GB — (6,326,705 × 46) float32
│   ├── train_labels.npy            51 MB  — (6,326,705,) int64
│   ├── train_anomaly_labels.npy    51 MB  — binary anomaly flag
│   ├── train_boundary_labels.npy   51 MB  — sequence boundary flag
│   ├── train_source_ids.npy        51 MB  — scenario/source ID per event
│   ├── val_* / test_*                     — same arrays for val and test splits
│   └── metadata.json                      — extraction config and statistics
│
├── loso_v3/                        Leave-One-Scenario-Out pre-split folds
│   ├── fold_1/ … fold_8/           One folder per held-out scenario
│   │   ├── train_features.npy
│   │   ├── train_labels.npy
│   │   ├── … (5 arrays × train/val/test splits)
│   │   └── metadata.json
│
├── augmented/                      Synthetic data used in paper training pipeline
│   └── tabddpm_v3/                 TabDDPM v3 (post-processed) — used in training
│       ├── events.parquet          170 KB — 2,328 synthetic events
│       │                           T1546/T1136/T1041/T1005/T1070 upsampled to 500/technique
│       │                           T1098 excluded (mode collapse)
│       └── metadata.json           generation config + post-processing steps
│
├── normal_only/                    Phase 3 pure normal traffic (pre-attack)
│   └── events.parquet              231 MB — 515,986 events, all is_attack=False
│                                   Collected March 6-7 2026, before any attacks ran.
│                                   Use for normal-signal pretraining / anomaly detection.
│
├── metadata/                       Dataset-level integrity files
│   ├── checksums.sha256            SHA-256 checksums for all primary data files
│   └── statistics.json             Per-split event counts, class distributions, etc.
│
└── val_versions/                   Validation index arrays
    ├── v1_indices.npy              T1110-capped validation indices
    ├── test_v1_indices.npy         T1110-capped test indices
    ├── v0_indices.npy              Uncapped validation indices
    └── v2_indices.npy              Alternative split indices

---

Classes (15)

Label index	MITRE ID	Technique	Train events (features_v3)
0	—	Normal	4,581,535
1	T1190	Exploit Public-Facing Application (SQL Injection)	473,215
2	T1110	Brute Force (capped at 50K in features)	50,000
3	T1078	Valid Accounts	1,270
4	T1059.004	Command and Scripting: Unix Shell (UDF exec)	361
5	T1136	Create Account	500
6	T1098	Account Manipulation	106
7	T1546	Event Triggered Execution (Trigger backdoor)	500
8	T1213	Data from Information Repositories	1,216,674
9	T1005	Data from Local System	500
10	T1029	Scheduled Transfer	566
11	T1041	Exfiltration Over C2 Channel	500
12	T1486	Data Encrypted for Impact (ransomware)	352
13	T1485	Data Destruction	126
14	T1070	Indicator Removal	500

Note: T1110 (Brute Force) has ~8.2M raw events in labeled/. The features_v3/ split caps T1110 at 50,000 to reduce class imbalance during training.

---

Data Splits

Standard split (features_v3/)

Split	Events	Description
Train	6,326,705	S01–S08 combined, T1110 capped at 50K
Validation	2,250,771	Held-out time window, T1110 capped (v1 indices)
Test	1,919,768	Separate held-out set (test_v1_indices)

Leave-One-Scenario-Out (loso_v3/)

8 folds, each holding out one complete attack scenario:

Fold	Held-out	Test attack techniques	Test events
1	S01	T1190, T1213	1,802,950
2	S02	T1110, T1078, T1098, T1213	10,227,922
3	S03	T1078, T1213, T1029	810,446
4	S04	T1190, T1098, T1486, T1485, T1070	808,394
5	S05	T1190, T1136, T1098, T1546	807,541
6	S06	T1190, T1213	1,737,982
7	S07	T1190, T1059.004, T1005	808,380
8	S08	T1190, T1059.004, T1213, T1041	821,660

---

Synthetic Data (augmented/)

Six rare-class techniques (T1546, T1136, T1041, T1005, T1070, T1098) have fewer than 110 original training events each. augmented/tabddpm_v3/ provides the synthetic oversampling used in the paper's training pipeline: 500 events per technique via TabDDPM (Kotelnikov et al., ICML 2023), with T1098 excluded due to mode collapse and constant-column artifacts corrected to match the train parquet schema.

See augmented/tabddpm_v3/metadata.json for the full post-processing log.

---

Feature Schema (46-dim float32)

Each log event is encoded as a 46-dimensional float32 vector.

Base features (dims 0–23, 24-dim)

Dims	ID	Feature	Encoding
[0:8]	Q1	query_type	One-hot: SELECT/INSERT/UPDATE/DELETE/CREATE/DROP/GRANT/OTHER
[8]	Q2	query_length	log1p(len(query))
[9]	Q5	has_union	Binary
[10]	Q6	has_subquery	Binary
[11]	Q7	accessed_tables_count	count / 10
[12]	Q10	has_system_table	Binary: mysql.* / information_schema
[13]	Q15	has_file_operation	Binary: LOAD DATA / INTO OUTFILE
[14]	Q17	has_privilege_command	Binary: GRANT/REVOKE/CREATE USER
[15]	Q18	has_user_management	Binary
[16]	S1	time_since_last	log1p(seconds)
[17]	S3	events_per_minute	count in last 60s / 100
[18:20]	T1	hour_of_day	sin + cos cyclical
[20]	T2	is_business_hours	Binary: 09:00–18:00
[21:23]	T4	day_of_week	sin + cos cyclical
[23]	U5	is_known_attacker_ip	Binary (reserved, currently 0)

Attack-signature features (dims 24–45, 22-dim)

Dims	ID	Feature	Encoding
[24]	A1	has_udf_function	Binary (T1059.004)
[25]	A2	has_encryption_func	Binary: AES_ENCRYPT/ENCODE (T1486)
[26]	A3	has_trigger_ddl	Binary: CREATE TRIGGER (T1546)
[27]	A4	has_log_manipulation	Binary: SET GLOBAL general_log (T1070)
[28]	A5	has_destructive_ddl	Binary: DROP TABLE/TRUNCATE (T1485)
[29]	A6	has_hex_payload	Binary: hex-encoded ELF/shellcode
[30]	A7	has_sleep_benchmark	Binary: SLEEP()/BENCHMARK()
[31]	A8	has_outfile	Binary: INTO OUTFILE/DUMPFILE
[32:43]	A9	event_type_onehot	11-way one-hot: QUERY/DDL/INSERT/UPDATE/DELETE/AUTH_FAILURE/CONNECT_FAILED/CONNECT/QUIT/SERVER_WARNING/SERVER_ERROR
[43]	A10	has_production_table	Binary
[44]	A11	query_keyword_count	count of attack keywords / 10
[45]	A12	has_concat_hex	Binary: CONCAT with hex strings

---

Parquet Schema (labeled/)

Each row represents a single MySQL log event:

Column	Type	Description
timestamp	datetime64[us, UTC]	Event timestamp
log_source	str	general, error, slow, or binlog
event_type	str	QUERY, DDL, INSERT, UPDATE, DELETE, AUTH_FAILURE, etc.
source_ip	str	Client IP (nullable)
user	str	MySQL user (nullable)
thread_id	int64	MySQL connection thread ID
db	str	Target database (nullable)
table	str	Target table (nullable)
query	str	SQL query text (nullable)
affected_rows	float64	Rows affected (nullable)
query_time	float64	Execution time in seconds (slow log)
lock_time	float64	Lock wait time in seconds (slow log)
rows_examined	float64	Rows examined (slow log)
rows_sent	float64	Rows returned to client
bytes_sent	float64	Response size in bytes
bytes_received	float64	Request size in bytes
error_code	str	MySQL error code (nullable)
severity	str	Log severity level (nullable)
raw_log	str	Original unparsed log line
log_file	str	Source log file path
log_position	float64	Byte offset in log file (binlog)
is_attack	bool	True if this event is part of an attack
scenario_id	str	S01–S08 (nullable for normal traffic)
run_id	str	Simulation run identifier
step	float64	Attack step index within scenario
mitre_technique	str	MITRE ATT&CK technique ID (null = Normal)
mitre_tactic	str	MITRE ATT&CK tactic
attack_path_position	str	start, middle, end, or isolated
confidence	float64	Label confidence (1.0 = verified)
group_id	str	Scenario group or collected for normal traffic

---

Log Source Distribution (train)

Source	Events	Description
general	6,884,967	All executed queries
error	6,658,525	Server events, auth failures, warnings
slow	930,509	Queries exceeding slow_query_log threshold
binlog	23,691	Binary log row-level change events

---

Attack Scenarios

Scenario	Primary techniques	Kill chain description
S01	T1190, T1213	SQL injection reconnaissance + data exfiltration
S02	T1110, T1078, T1098, T1213	Brute force → valid account → privilege escalation → data collection
S03	T1078, T1213, T1029	Valid account abuse → bulk data collection → scheduled transfer
S04	T1190, T1098, T1486, T1485, T1070	Injection → account manipulation → ransomware → data destruction → cover tracks
S05	T1190, T1136, T1098, T1546	Injection → backdoor account creation → trigger implant
S06	T1190, T1213	Large-scale SQL injection + bulk data theft
S07	T1190, T1059.004, T1005	Injection → UDF shellcode execution → local file read
S08	T1190, T1059.004, T1213, T1041	Injection → UDF → data collection → exfiltration over C2

---

Usage with the Paper Code

See the companion GitHub repository:

GitHub: https://github.com/Hyeonseop-Shin/DB-forensic-hybrid

git clone https://github.com/Hyeonseop-Shin/DB-forensic-hybrid
cd DB-forensic-hybrid

# Install package
SETUPTOOLS_USE_DISTUTILS=local pip install -e .

# Download pre-extracted features (2.2 GB, no GPU needed for evaluation)
python -c "
from huggingface_hub import snapshot_download
snapshot_download(
    repo_id='golfoscar/db-forensic-hybrid',
    repo_type='dataset',
    local_dir='data/',
    allow_patterns=['features_v3/*', 'val_versions/*'],
    token='YOUR_HF_TOKEN',
)
"

# Run test set evaluation with pre-trained checkpoints
python scripts/evaluate_test.py --config configs/E11.yaml --gpu 0

# Run LOSO evaluation
python scripts/evaluate_loso.py --gpu 0

# Re-extract features from raw logs (requires labeled/ + augmented/tabddpm_v3/, ~1 hour)
python scripts/extract_features.py --n-features 46

Re-running augmentation from scratch

# Download labeled train split and generate synthetic data
python scripts/generate_synthetic.py --model tabddpm --gpu 0
# Output: data/augmented/tabddpm_v3/events.parquet (applied automatically by extract_features.py)

---

Citation

@dataset{shin2026dbforensic,
  author    = {Shin, Hyeonseop},
  title     = {{DB-Forensic-Hybrid}: A Large-Scale {MySQL} Log Dataset
               for Event-Level Attack Detection and Attribution},
  year      = {2026},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/datasets/golfoscar/db-forensic-hybrid},
}

---

License

Released under CC BY 4.0. See DATASHEET.md for the full dataset card.