TeleAntiFraud-28k

TeleAntiFraud-28k is the first open-source audio-text slow-thinking dataset specifically designed for automated telecom fraud analysis. This dataset integrates audio signals with reasoning-oriented textual analysis, providing high-quality multimodal training data for telecom fraud detection research.

Dataset Overview

• Total Samples: 28,511 rigorously processed speech-text pairs
• Total Audio Duration: 307 hours
• Unique Feature: Detailed annotations for fraud reasoning
• Task Categories: Scenario classification, fraud detection, fraud type classification

Dataset Construction Strategies

1. Privacy-preserved Text-Truth Sample Generation

• Using ASR-transcribed call recordings (with anonymized original audio)
• Ensuring real-world consistency through TTS model regeneration
• Strict adherence to privacy protection standards

2. Semantic Enhancement

• LLM-based self-instruction sampling on authentic ASR outputs
• Expanding scenario coverage to improve model generalization
• Enriching the diversity of conversational contexts

3. Multi-agent Adversarial Synthesis

• Simulation of emerging fraud tactics
• Generation through predefined communication scenarios and fraud typologies
• Enhancing dataset adaptability to new fraud techniques

TeleAntiFraud-Bench

We have constructed TeleAntiFraud-Bench, a standardized evaluation benchmark comprising proportionally sampled instances from TeleAntiFraud-28k, to facilitate systematic testing of model performance and reasoning capabilities on telecom fraud detection tasks.

Model Contribution

We contribute a production-optimized supervised fine-tuning (SFT) model based on Qwen2-Audio, trained on the TeleAntiFraud training set.

Examples

Explore our dataset examples to better understand the telecom fraud detection capabilities:

• Case 1: Normal Conversation Analysis - Detailed analysis of a legitimate phone conversation
• Case 2: Fraud Conversation Analysis - Step-by-step reasoning for detecting a fraudulent call
• Evaluation Sample - Representative sample from our evaluation benchmark
• Model Output: Normal Conversation - Our model's reasoning process on a legitimate call
• Model Output: Fraud Detection - Model's analysis and detection of a fraudulent call

Multi-Agent Data Collection

To collect fraudulent conversation data:

1. Insert your API key in multi-agents-tools/AntiFraudMatrix/main.py (uses SiliconFlow API key)
2. Run the following command to generate fraudulent dialog text:

   python multi-agents-tools/AntiFraudMatrix/main.py

3. Results will be saved in the result directory

For normal conversation data:

• Use multi-agents-tools/AntiFraudMatrix-normal/main.py following the same process

Voice Synthesis with ChatTTS

To synthesize speech from the collected text:

1. Install the necessary dependencies

2. Run the API server:

   fastapi dev ChatTTS/examples/api/main_new_new.py --host 0.0.0.0 --port 8006

3. Use any of the scripts in ChatTTS/examples/api/normal_run*.sh or ChatTTS/examples/api/run*.sh

   Modify the port in these scripts if needed, then run:

   bash ChatTTS/examples/api/run*.sh

Open-Source Resources

• TeleAntiFraud-28k dataset
• TeleAntiFraud-Bench evaluation benchmark
• Data processing framework (supporting community-driven dataset expansion)
• TeleAntiFraud-Qwen2-Audio SFT model

Key Contributions

1. Establishing a foundational framework for multimodal anti-fraud research
2. Addressing critical challenges in data privacy and scenario diversity
3. Providing high-quality training data for telecom fraud detection
4. Open-sourcing data processing tools to enable community collaboration

Citation

@inproceedings{Ma2025TeleAntiFraud28kAA,
  title={TeleAntiFraud-28k: An Audio-Text Slow-Thinking Dataset for Telecom Fraud Detection},
  author={Zhiming Ma and Peidong Wang and Minhua Huang and Jingpeng Wang and Kai Wu and Xiangzhao Lv and Yachun Pang and Yin Yang and Wenjie Tang and Yuchen Kang},
  year={2025},
  url={https://api.semanticscholar.org/CorpusID:277467703}
}