Hugging Face's logo

Datasets:
rypow
/
dendroaspis-tetragon-hids

Tasks:
Tabular Classification
Other
Modalities:
Tabular
Text
Formats:
parquet
Languages:
English
Size:
10M - 100M
Tags:
cybersecurity
host-intrusion-detection
anomaly-detection
tetragon
ebpf
sequence-modeling
Libraries:
Datasets
Dask
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
dendroaspis-tetragon-hids / datacard.md
rypow's picture
rypow
v0.2.1 - clean release
d41f3fb verified
preview code
|
raw
history blame
19.5 kB
metadata 
license: cc-by-4.0
task_categories:
  - tabular-classification
  - other
tags:
  - host-intrusion-detection
  - tetragon
  - ebpf
  - sequence-modeling
  - att-and-ck
  - anomaly-detection
size_categories:
  - 10M<n<100M
language: []
pretty_name: v0.2 Tetragon-native HIDS corpus (22 technique IDs)

Dataset summary

  • Source: Tetragon 1.6.1 eBPF telemetry on a single LXC container, 6-day collection window 2026-04-14 → 2026-04-20.
  • Population: all process / file / network / credential / memory events attributable to the target container's PID-namespace inum across the collection window (parser filters host-namespace events at parse time).
  • Splits: chronological — train (2026-04-14T01:41Z → 2026-04-19T22:09Z) / test (2026-04-19T21:46Z → 2026-04-20T02:01Z). All ART attack intervals fall in the test split; train is benign-only by construction.
  • Format: Parquet shards under data/, suitable for the Hugging Face datasets library streaming loader. Labels are a separate labels.csv.
  • License: CC BY 4.0.

Supported tasks

  • Host-telemetry anomaly detection. Per-event or per-window scoring against the time-window labels in labels.csv.
  • Sequence anomaly modeling. Autoregressive next-event prediction (NLL) or masked-event modeling (MEM, including field-aware MEM-FA) over the integer-coded feature columns.
  • Event-stream classification baselines. N-gram, IsolationForest, XGBoost, etc. operate directly on the categorical / hash columns.

The corpus is not intended for: attack-payload synthesis, ground-truth event-level labels (only intervals are labeled), or production deployment without recalibration to the operator's own host distribution.

Dataset structure

Splits

Split Rows Time span (UTC)† Attack events Attack rate
train 16,951,950 2026-04-14 01:41:13 → 2026-04-19 22:09:18 0 0.0 %
test 2,349,965 2026-04-19 21:46:27 → 2026-04-20 02:01:27 705,852 30.0 %

† Pre-window event timestamps. ≈295 events (267 train + 28 test, 0.0015 % of the corpus) carry event_time values that predate the stated collection window. All have f_is_procfs_walk = 1: parser-synthesized process_exec records for daemons (systemd, bash, dbus-daemon, polkitd, cron, sshd, etc.) that were already running when Tetragon collection started, using the daemon's original process_start_time as the event timestamp. They preserve lineage roots for downstream events. Filter WHERE f_is_procfs_walk = 0 to exclude them from time-window analysis.

Schema evolution (v0.2.1)

The release artifact is the encoded parquet (63 columns: 39 model-input columns + 7 hash columns + identifier and side-string columns). Compared to v0.2.0 the encoded schema gained four 3a′ promoted categoricals (f_path_category, f_dst_ip_category, f_dst_port_category, f_object_category) and lost two constant-zero booleans (f_in_init_tree, f_args_truncated); two hash columns were right-sized (f_proc_name_hash 1024 → 2048, f_proc_cwd_hash 256 → 4096 with depth 2 → 4 path components hashed). The full integer-code → label mapping is in reproduction/feature_definitions.json (schema version 1.1).

The release pipeline is three-stage:

Stage Source field Stage column Transformation
Raw Tetragon JSON process_exec / process_exit / process_kprobe envelope event_type passthrough → vocab f_event_type
→ Parser (126-col raw, withheld) process.binary proc_binary (full path) derive basename → proc_binary_basename (released, aliased) + hash f_proc_name_hash
process.cwd proc_cwd (full path) normalize → proc_cwd_normalized (path-rewritten at release) + hash f_proc_cwd_hash
process.uid proc_uid (uint32) bucket → f_proc_uid_bucket (uint8) — raw uid not released
process.arguments proc_arguments length-bucket → f_args_length_bucket (uint8); raw args not released
process_kprobe.args[].sock.daddr kp_sock_daddr category bucket → dst_ip_category (string) — raw IP not released
→ Behavior builder (encoded, released) (event_type, kprobe_function_name) f_action_family (uint8, 15 codes) derived; see feature_definitions.json
(parent_basename, proc_basename) pair f_parent_child_pair_hash (uint16, 1024 buckets) hash
proc_exec_id walk f_lineage_depth, f_root_ancestor_basename_hash, f_process_tree_id_hash derived

The 126-column raw parser output is withheld from the public release. All raw paths, raw arguments, raw IPs, raw uids, hostnames, and PID-namespace inums are dropped at this stage; see "Personal and sensitive information" below for the per-column decisions.

Feature column inventory (encoded parquet)

Full integer-code → label mapping for every categorical column lives in reproduction/feature_definitions.json. High-level summary:

  • 22 integer-coded categorical columns — 18 from v0.2.0 plus the four 3a′ promoted categoricals (f_path_category 11-way, f_dst_ip_category 6-way, f_dst_port_category 7-way, f_object_category 8-way).
  • 4 string-coded categorical columns (path_category, object_category, dst_ip_category, dst_port_category) — provenance side columns, identical semantics to their f_* integer counterparts.
  • 7 hash columns with declared bucket capacities (2048 / 1024 / 4096 / 256 / 1024 / 1024 / 4096 — see Hash-bucket utilization below).
  • 6 boolean flag columns (e.g. f_is_procfs_walk, f_uid_eq_parent, f_kp_commit_creds_uid_change).
  • 5 numeric columns (delta_t_prev_ns, process_age_ns, mmap/mprotect prot bitmasks, fd integer).
  • 3 bucketed log-time columns (f_delta_t_log_bucket, f_process_age_log_bucket, f_lineage_depth).
  • 3 identifier columns (event_time, proc_exec_id, proc_pid).
  • 7 release-eligible side-string columns (basenames, normalized cwd, parent-child pair, token, process_tree_root_exec_id).

Class distributions (train split)

f_event_type (4-way; OOV not observed):

Code Label Count %
2 process_kprobe 16,413,759 96.83
1 process_exit 288,716 1.70
0 process_exec 249,475 1.47

f_action_family (15-way; top 10 shown; full table in feature_definitions.json):

Code Label Count
6 FILE_OPEN 14,852,286
12 MEM_MAP 1,168,402
1 PROC_EXIT 288,716
9 PRIV_CHANGE 257,709
0 PROC_EXEC 249,475
7 FILE_DELETE 44,818
5 NET_SEND 23,672
13 MEM_PROTECT 19,059
3 NET_CLOSE 18,451
2 NET_CONNECT 16,359

ATT&CK label coverage

61 labeled time intervals span 21 ATT&CK techniques + 1 supply-chain capstone chain across 8 tactics. All attack-positive intervals are in the test split. Per-technique event counts (test split, top 10 by event volume):

ID Technique Test events
T1059.004 Unix Shell 175,797
T1083 File and Directory Discovery 154,581
T1552.001 Credentials In Files 97,493
T1543.002 Systemd Service 47,901
CHAIN.CAP01 Capstone Chain v0.1 (Supply-Chain Narrative) 34,130
T1548.003 Sudo and Sudo Caching 24,229
T1059.006 Python 22,675
T1562.001 Disable or Modify Tools 18,610
T1195.002 Compromise Software Supply Chain 15,703
T1548.001 Setuid and Setgid 12,461

Full per-technique table is reproducible from notebooks/v0-2_dataset_audit.ipynb (cell 63).

Hash-bucket utilization (train split, post-3a′ caps)

Column Capacity Buckets used Saturation
f_proc_name_hash 2048 269 13.1 %
f_parent_proc_hash 1024 121 11.8 %
f_proc_cwd_hash 4096 65 1.6 %
f_lineage_bag_hash 256 98 38.3 %
f_parent_child_pair_hash 1024 490 47.9 %
f_root_ancestor_basename_hash 1024 5 0.5 %
f_process_tree_id_hash 4096 7 0.2 %

Low saturation reflects the single-host single-developer baseline (a small real basename / cwd / lineage population). The 3a′ capacity bumps on f_proc_name_hash and f_proc_cwd_hash were sized for plausible multi-host saturation rather than this corpus's observed counts; the values are documented here so re-trainers can resize the embedding tables.

Dataset creation

Curation rationale

This dataset was assembled to support the v0.2 milestone of an academic project investigating Mamba-based anomaly detection on host telemetry. The v0.1 predecessor produced an inverted AUROC (~0.35) on a similar but parser-buggy corpus; v0.2 was rebuilt with a Tetragon-native parser, sentinel-window exclusion at parse time, and a two-stage parser → behavior-builder design that preserves a raw normalized parquet internally for iteration. The v0.2.1 release further re-parsed the same collection on 2026-04-29 with five correctness fixes:

  • Procfs-walk synthetic execs are now tagged, not dropped (1,346 events recovered corpus-wide; lineage gaps for daemons running before agent start closed).
  • delta_t_prev_ns is keyed per-process rather than globally, removing 44,947 train + 1,643 test negative residuals.
  • Test-split lineage walking now seeds from the train parquet first, resolving 39 % of previously-missing parents (test lineage_missing_parent: 5.25 % → 2.88 %).
  • Parser parallelized across files via multiprocessing.Pool, with DuckDB ORDER BY event_time for chronological output.
  • Mamba NLL scorer's front-pad alignment fix: nll_padded[i] is now event i's surprise (was i+1's in v0.2.0).

Source data

  • Sensor: Tetragon 1.6.1 (Cilium eBPF) running on a Linux host.
  • Workload: a single LXC container hosting a developer's daily workspace (Node, Python, VSCode, npm). Background services inside the container (systemd-resolved, dbus, polkitd, etc.) contribute system noise.
  • Tracing policies: 7 YAML policies covering file, network, memory, privilege, process-injection, and file-attribute kprobes (bundled in reproduction/tetragon-configs/).
  • Attribution: the parser pins on PID-namespace inum, which uniquely identifies the target container across an inum rotation that occurred during the collection window. Host-namespace events are dropped at parse time.

Annotations

Time-window labels are provided in labels.csv (61 rows; 22 distinct technique IDs incl. one capstone chain; 8 tactics). Labels were generated by an Atomic Red Team campaign run inside the same container during the test window, with start/end timestamps captured per-trial. The campaign manifest is bundled at reproduction/agentic-art-campaign/.

Personal and sensitive information

component-first anonymization with deep aliasing across all PII-bearing string columns. Summary:

  • 126-column raw parquet withheld. All raw paths, raw arguments, raw IPs, raw uids, hostnames, and PID-namespace inums are dropped at the parser stage. The released parquet is the 63-column encoded form.
  • Basename aliasing. A single alias map covers proc_binary_basename, parent_binary_basename, and root_ancestor_basename. Distro-standard binaries and dev-tooling that anchor the "standardized agentic-research workstation" baseline (bash, node, python3.12, claude, gh, npm, code, git, etc.) pass through verbatim. Custom binaries (ART trial scripts, build artifacts) become custom_tool_NNN (lex-sorted). VSCode commit-hash binaries (code-<40hex>) collapse to a single vscode_server_bin alias. The full original → alias map is shipped at reproduction/alias_map.json.
  • Composite re-composition. parent_child_pair and token are regenerated from the aliased atomic columns via the canonical format_token, so the original (un-aliased) basename never re-enters these fields.
  • Path rewriting. proc_cwd_normalized is rewritten by ordered rules: /home/<user> → /home/_user_, /Users/<user> → /Users/_user_, /mnt/projects/<X> → /workspace/_project_, /projects/<X> → /workspace/_project_, residual rypow_ → custom_.
  • Exec-id pseudonymization. proc_exec_id and process_tree_root_exec_id are rewritten as exec_<32hex> via salted HMAC-SHA256 (16-byte digest, hex-encoded). Equality joins are preserved (lineage walks still work); the original base64-encoded host:ts:pid payload is erased.
  • Defensive global string sweep. A final substring pass over every string column substitutes rypow → research_user, prox1 → sandbox_host, dendroaspis → agent_platform_v02, ryan_admin → privileged_user as a safety net.
  • Post-write assertions. A regex sweep over the released shards (notebook §22-24) confirms zero matches on raw or base64-decoded values for any of: prox1, rypow, ryan_admin, /home/rypow/, /projects/ml_final_project, /mnt/projects/, dendroaspis, /datasets/, MAC-address pattern. proc_exec_id and process_tree_root_exec_id are format-checked against ^exec_[0-9a-f]{32}$. Every basename in the three basename columns is asserted to be in the alias-map closure.

Considerations for using the data

Sources of bias

  • Single host, single developer. All baseline (benign) telemetry comes from one Linux container running one developer's daily workflow over six days. Tool habits (VSCode, npm, git, claude-code) appear at high frequency; absent activities (CI/CD pipelines, multi-tenant traffic, GPU jobs) are simply not represented.
  • Container-internal UIDs. Tetragon 1.6.1 emits container-internal UIDs; the encoder buckets uids into 5 classes (ROOT / RESEARCH_USER / SYSTEM_LOW / DAEMON_HIGH / OTHER) rather than treating them as a continuous identifier. Researchers reusing these buckets on multi-host data should re-bucket.
  • Time skew. Train spans 5 d 20 h, test spans only 4 h 15 m. The test window concentrates 30 % attack-event mass, which is unrealistically dense for production traffic; downstream metrics (AUROC, AP) on this corpus reflect the campaign's deliberate density.
  • Hash-bucket sparsity. Hash columns are sized for plausible multi-host saturation; on this single-host corpus most are < 50 % utilized (see hash-bucket table above).
  • Aliased basename narrative. The v0.2.1 anonymizer collapses a long tail of one-off custom binaries into custom_tool_NNN slots and folds per-build VSCode commit-hash binaries into vscode_server_bin. The resulting parquet looks like a "standardized agentic-research workstation" baseline more than the specific developer's machine; the true diversity of custom-script names was higher than what survives in the released string columns.

Social impact

This is a defensive-research dataset. It enables researchers to develop and benchmark host-anomaly-detection models without operating live attack infrastructure. The bundled ATT&CK technique inventory is not an attack tutorial — payload generation requires a separate Atomic Red Team installation and is out of scope.

Known limitations

  • No ground-truth event-level labels. Labels are time-window only; some benign system activity (e.g. background daemons running during an ART technique window) is labeled-attack by association.
  • Train has zero attack events. All ART intervals fall in test by construction. Unsupervised / self-supervised methods are the natural fit; fully supervised classifiers would need to engineer their own train-time positives.
  • Single Tetragon version. Format quirks of Tetragon 1.6.1 (e.g. KPROBE_ACTION enum coverage, ns-PID inum semantics) are baked in. Older / newer Tetragon emits slightly different envelopes; the parser pin in reproduction/PARSER_VERSION documents the exact reference build.
  • Linux only. Tetragon is Linux-only; this corpus has no Windows or macOS analogue.

Intended use + dual-use disclaimer

Intended use. This dataset supports:

  • training and evaluating sequence / anomaly models on host telemetry,
  • benchmarking baselines (n-gram, XGBoost, IsolationForest, LSTM, Mamba, Transformer) on a fixed split,
  • studying ATT&CK-technique-specific detectability,
  • research on attack representations, lineage features, and event encoders.

Not intended use. The attack telemetry is real ATT&CK technique execution but generated in an isolated lab LXC container; the corpus is not an attack template, not a payload library, and not a substitute for production EDR data. Operators who reuse the bundled tracing policies + parser to collect their own corpus are responsible for their own sandboxing.

Additional information

License

CC BY 4.0 — see LICENSE in this repo. Code under reproduction/scripts/ (in the project repository) is MIT-licensed.

Citation 

@dataset{dendroaspis_v0_2_2026,
  title   = {v0.2 Tetragon-native HIDS corpus with 21 ATT\&CK techniques},
  author  = {Powers, Ryan William},
  year    = {2026},
  version = {0.2.1},
  doi     = {<TBD-on-publish>},
  url     = {<TBD-on-publish>}
}

Contact

GitHub: @ryypow. Issues / corrections preferred via the project repository.

repos: https://github.com/ryypow/dendroaspis