FraudRecon

FraudRecon is a Chinese audio-text dataset for telecom-fraud understanding. It contains case-grounded synthetic multi-turn dialogues, paired synthetic speech, and hierarchical annotations for fraud and matched benign interactions.

This repository describes release v1.1. Complete archives are distributed through Hugging Face gated access. This README is publicly visible. Access to the dataset files, including archives, metadata, and supporting documentation, requires approval.

Data construction

Fraud dialogues are produced with Narrative-to-Dialogue Reverse Reconstruction (NDRR), which converts de-identified public case narratives into grounded procedural representations and then synthesizes dialogue turns under those constraints. Matched benign dialogues are constructed from de-identified service dialogues and templates. The resulting text is rendered with TTS and telephone-channel degradation.

Release summary

Split	Total	Fraud	Benign, stored as normal
train	72,112	37,935	34,177
test	8,207	4,213	3,994
Total	80,319	42,148	38,171

Additional properties:

• Language: Chinese.
• Modalities: structured dialogue text and paired MP3 audio.
• Audio: 48 kHz, stereo, nominally 128 kbps.
• Taxonomy: 12 fraud categories, 12 matched benign categories, 6 fraud phase classes, 3 benign phase classes, 26 fraud-tactic signals, and 16 benign service-behavior signals.
• Sample identifiers: case_id uniquely identifies each sample.

The stored value normal denotes benign (non-fraud) dialogues.

Source and generation overview

The source pool contains more than 50,000 publicly available fraud narratives, including victim posts, news reports, and police safety materials. Source narratives were de-identified before reconstruction.

NDRR consists of:

1. Fraud Procedure Extraction, which converts a narrative into a case-grounded procedural representation.
2. Fact-Anchored Dialogue Reconstruction, which synthesizes phase-structured scammer-victim turns under the extracted constraints.
3. Consistency Verification, which checks factual consistency, procedural fidelity, and dialogue naturalness.

For training-set balancing, 3,986 additional tail-category fraud dialogues are generated from category-level priors over phase transitions, interaction signals, and role profiles. These samples are included in the training split. The benign training distribution is matched accordingly.

Audio is synthesized with ChatTTS and processed to simulate telephone-channel characteristics for multimodal benchmarking.

Supported benchmark settings

• T1 Fraud Detection: predict whether a dialogue is fraudulent.
• T2 Category Classification: predict the Level-1 fraud or matched benign category.
• T3 Phase Identification: identify the procedural phase of a dialogue segment.
• T4 Interaction Signal Recognition: predict applicable fraud tactics or benign service behaviors.
• T5 Evidence-Chain Reasoning: produce a four-step chain connecting entities, phases, signals, and final judgment.

For evaluation, use dialogue text and/or audio as model inputs and the corresponding annotations as prediction targets. See Schema and evaluation views for field details.

Annotation resources

• Annotation codebook: definitions for every released category, phase, and signal.
• Schema and evaluation views: field types, conditional train/test fields, path rules, and the model-input whitelist.
• metadata/category_distribution.csv: complete train/test category counts.
• metadata/test/test_labels.jsonl: test gold annotations.

phase_id is a local phase-instance identifier. Use phase_type for the canonical phase class. Phases may be omitted or repeated according to the case-specific trajectory.

v1.1 archive organization

v1.1 contains 20 ZIP64 archives using DEFLATE level 6, organized by split and data_type:

• Train fraud: 12 archives containing 2,100 to 3,700 samples; exact counts and case ranges are recorded in metadata/shard_manifest.csv.
• Train normal: 5 archives containing 6,300 to 8,700 samples; exact counts and case ranges are recorded in metadata/shard_manifest.csv.
• Test fraud: 2 archives containing 2,100 and 2,113 samples.
• Test normal: 1 archive containing all 3,994 samples.

Every archive contains exactly one data_type: fraud or normal. It also contains both the JSON and MP3 for every assigned case. Each archive covers a contiguous numeric case_id interval. Boundaries are selected in 100-sample steps to balance compressed sizes. The final archive in each split and data_type group contains the remaining samples.

FraudRecon/
  README.md
  LICENSE
  dataset_manifest.json
  release_notes.md
  docs/
    CODEBOOK.md
    SCHEMA.md
  metadata/
    shard_manifest.csv
    sample_to_shard.jsonl
    checksums.sha256
    train/
    test/
  preview/
    train/
      metadata.csv
      samples/
      audio/
    test/
      metadata.csv
      samples/
      audio/
  archives/
    train/
      FraudRecon-train-001-v1.1.zip
      ...
      FraudRecon-train-017-v1.1.zip
    test/
      FraudRecon-test-001-v1.1.zip
      ...
      FraudRecon-test-003-v1.1.zip

Train files 001-012 contain fraud samples and 013-017 contain normal samples. Test files 001-002 contain fraud samples and 003 contains normal samples. metadata/shard_manifest.csv is the authoritative shard table and records exact sample counts, first and last case_id, sizes, and checksums. metadata/sample_to_shard.jsonl provides exact case-to-archive lookup.

Hugging Face preview

The default preview configuration contains 24 training and 24 test examples, with paired audio, speaker-tagged transcriptions, and descriptive metadata. The complete dataset is provided in the ZIP archives.

The preview uses the MP3 files in preview/train/audio/ and preview/test/audio/, together with:

• preview/train/metadata.csv
• preview/test/metadata.csv

Both files use the same CSV schema with file_name and transcription columns. Audio paths are relative to the metadata CSV directory.

Extracting and loading

• Sample files: <case_id>.json. Case IDs use the prefixes train_fraud_, train_normal_, test_fraud_, and test_normal_, for example train_fraud_00001 or test_normal_00001.
• Audio files: <case_id>_mix.mp3, paired with the JSON file for the same case.

Each archive preserves split-relative paths:

train/
  samples/
    train_fraud_00001.json
  audio/
    train_fraud_00001_mix.mp3

Extract archives from the same split into a shared directory. Resolve each sample's audio_path relative to its extracted train/ or test/ directory.

Train labels are stored inside sample JSON files. Test labels are stored in metadata/test/test_labels.jsonl.

Integrity verification

metadata/shard_manifest.csv records actual archive bytes, member counts, compression ratios, and SHA-256 values. For a complete download, verify all archives from the repository root:

sha256sum -c metadata/checksums.sha256

A complete release contains:

• 20 archives, each containing exactly one data_type.
• 80,319 unique case IDs.
• 80,319 JSON files.
• 80,319 paired MP3 files.

Access

Dataset files are available after approval of a Hugging Face access request. Applicants should provide an institutional affiliation and describe their intended research use. Access administration is separate from the copyright license; approved users must follow the LICENSE file and any terms displayed in the Hugging Face access request.

Privacy

Source narratives undergo automated de-identification and manual spot checks. Numeric identifiers in the generated dialogues, including phone numbers, account numbers, identity numbers, order numbers, and transaction identifiers, were reviewed and confirmed to be synthetic.

Limitations

• Synthetic dialogues and TTS audio have limited coverage of real-world interaction patterns and acoustic variability.
• Coverage reflects the Chinese source pool, with possible topical and reporting biases from public narratives and service templates.
• The test split is category-stratified; class proportions reflect benchmark sampling.
• Automated data construction and annotation may retain model biases and annotation-rubric assumptions.

Results on FraudRecon may differ from performance in real-world scenarios.

Responsible use

FraudRecon is intended for research on fraud understanding, risk detection, dialogue analysis, safety, and audio-text model evaluation. It should not be used to facilitate impersonation, coercion, evasion, or operational fraud.

Citation

If you use FraudRecon, cite:

@inproceedings{liu2026fraudrecon,
  title     = {Every Victim Tells a Dialogue: Reconstructing 80K
               Audio-Text Anti-Fraud Conversations from Case Narratives},
  author    = {Jingxuan Liu and Haoyang He and Xuan Yang and Shuyuan Zhu
               and Ke Li and Hong Wu and Lan Yang and Jian Chang
               and Honggang Zhang},
  booktitle = {Proceedings of the 34th ACM International Conference on
               Multimedia},
  year      = {2026},
  doi       = {10.1145/3767308.3835796}
}

License

FraudRecon is licensed under the Creative Commons Attribution 4.0 International License. See LICENSE. The gated-access workflow controls initial access; reuse after approval is governed by the applicable license and access terms.

Version history

See release notes for v1.1 updates.