--- pretty_name: IV-CAN-v1 license: cc-by-4.0 task_categories: - tabular-classification tags: - automotive - can-bus - cybersecurity - intrusion-detection - time-series size_categories: - 100M/{data.csv,label.yaml} │ └── Attack///{data.csv,label.yaml} ├── analysis/ # case-level plots and tracing artifacts ├── code/attack_generation/ # self-contained reproducibility snapshot ├── schema/ # machine-readable schemas ├── statistics/ # aggregate statistics and figures ├── tools/validate_dataset.py # lightweight public validator ├── dataset_manifest.csv ├── LICENSE.md ├── SCHEMA.md └── README.md ``` Each case has a stable public id such as `fuzzing_0001`. The manifest provides all public data and analysis paths. Large CSV files are intentionally not rendered by the Hub viewer; stream them or download selected cases instead. ## Data schema Every `data.csv` uses this ordered header: ```text timestamp,can_id,is_extended_id,dlc,data,RX_or_TX,type,is_fd,domain,is_attack,traced_from,uuid,is_traced_attack ``` The normative definitions, units, enums, identifier scope, and tracing constraints are in [SCHEMA.md](SCHEMA.md). Machine-readable definitions are provided in [schema/frame.schema.yaml](schema/frame.schema.yaml) and [schema/label.schema.json](schema/label.schema.json). Important interpretation rules: - `timestamp` is Unix epoch time in seconds, expressed on the UTC time scale. The original recorder wall-clock was interpreted in the project acquisition timezone (`Asia/Shanghai`) during conversion. - `RX` and `TX` are relative to the recording channel: received from or transmitted by that channel. - PEAK frame types are `DT` (data frame without the combined flags represented below), `FB` (CAN FD with BRS), and `BI` (CAN FD with BRS and ESI). Use `is_fd`, not `DT` alone, to distinguish classical CAN from CAN FD. - `uuid` is unique only within one case. `traced_from` contains zero or more parent UUIDs from that same case; UUIDs must never be joined across cases. - `is_attack=true` identifies injected attack frames. A forwarded-only attack-related frame is selected by `is_traced_attack=true AND is_attack=false`; the flags are not declared mutually exclusive. Run a bounded structural check from the repository root: ```bash python tools/validate_dataset.py --max-rows 1000 ``` Use `--all` to scan every row (this is expensive for the full dataset). ## Labels and trial categories | Category | Trials | | --- | ---: | | Benign | 60 | | DoS | 22 | | fuzzing | 28 | | replay | 12 | | spoofing | 8 | | suspension | 2 | | masquerade | 2 | Public attack-type names are case-sensitive and normalized exactly as shown. Case ids use lowercase prefixes, for example `dos_0001` and `replay_0001`. ## Attack-generation code [code/attack_generation](code/attack_generation/) is a curated, self-contained snapshot of the core attack construction logic used for the dataset. It excludes private builder infrastructure and third-party hardware binaries. DoS, fuzzing, replay, and spoofing include frame-generation/execution paths; suspension and masquerade are published as validated experiment plans because their execution requires a two-interface MitM bridge. The command-line tool defaults to a no-transmit preview. Hardware execution requires two explicit safety flags and an authorized, isolated CAN test bench. See the [code README](code/attack_generation/README.md) for installation, limitations, and tests. ## Collection and attack statistics | Metric | Value | | --- | ---: | | Total trials | 134 | | Total CAN frames | 225,850,423 | | Trials with non-driver control | 67 (37 benign; 30 attack) | | Attack-execution trial duration (primary) | 3h9min | | Cumulative injection-window duration (supplementary) | 23min | | `is_attack` frames | 231,477 | | `is_traced_attack` frames | 42,478 | | Trials with attack cross-domain forwarding | 23 | | Domain | Configured name | Protocol | Frames | | --- | --- | --- | ---: | | 1 | 驱动CAN | CAN FD | 57,786,350 | | 2 | 底盘CAN | CAN FD | 64,904,112 | | 3 | 灯控CAN | Classical CAN | 48,815,343 | | 4 | 辅助行驶CAN | CAN FD | 47,144,853 | | 5 | 方向盘CAN | Classical CAN | 7,199,765 | The primary **attack-execution trial duration** is the sum of the full `collection_duration_sec` spans of all 74 attack trials (11,388.886 seconds, reported as 3h9min). It represents the complete driving and observation context in which attacks were executed. The supplementary **cumulative injection-window duration** sums `attacking_duration_sec` across those trials (1,429.850 seconds, reported as 23min); it covers only the configured active injection windows and excludes the surrounding observation periods. The attack-effect summary is available in [statistics/attack_effect_counter.md](statistics/attack_effect_counter.md), with additional machine-readable statistics in [statistics/count_csv_result.yaml](statistics/count_csv_result.yaml). ## Responsible use This dataset and supplementary code are intended for authorized security research, reproducibility, defensive evaluation, and education. Do not use attack-generation functionality on public roads, operational vehicles, or networks you do not own or have explicit permission to test. ## License and attribution IV-CAN-v1 data, original documentation, schemas, figures, and the original supplementary attack-generation code are released under [CC BY 4.0](LICENSE.md). Attribution is required. Third-party drivers and runtimes are not redistributed and remain subject to their vendors' licenses.