CiferAI/cifer-fraud-detection-k1-a

🧠 Cifer Fraud Detection Model

(cifer-fraud-detection-k1-a)

🧾 Overview

This model is a binary classifier trained to detect fraudulent transactions using the Cifer Fraud Detection Dataset (6 million synthetic rows). It is designed to operate in federated learning environments, where data is split across clients or organizations without centralized access.

This model was trained on 6 million synthetic rows, split into four partitions of 1.5 million records each. You can train this model independently across the four dataset partitions, then aggregate the results using FedAvg (Federated Averaging) to achieve performance comparable to centralized training—as validated in Cifer’s internal lab benchmarks.

This model is part of Cifer’s laboratory-validated framework for privacy-preserving machine learning, enabling secure, consent-first collaboration without exposing raw data. It is fully compatible with Cifer’s no-code workspace and federated orchestration engine.

---

📊 Training Data

• Dataset: CiferAI/Cifer-Fraud-Detection-Dataset-AF
• Total rows: 6,000,000 (split into 4 federated parts)
• Type: Fully synthetic tabular data modeled after real-world financial fraud scenarios
• Fields: transaction type, amount, sender/receiver balance, fraud flags, and step-based timestamps
• Generated with: Cifer Simulation Engine, modeled after the PaySim simulator

---

🧠 Model Architecture

• Framework: TensorFlow / Keras
• Architecture: Multi-layer Perceptron (MLP)
• Layers:
  • Input Layer (shape = number of features)
  • Dense(64, activation="relu")
  • Dense(32, activation="relu")
  • Dense(2, activation="softmax")
• Loss Function: sparse_categorical_crossentropy
• Optimizer: adam
• Output: Fraud probability classification (0 = normal, 1 = fraud)

---

✅ Use Cases

• Fraud detection in fintech, mobile money, and digital banking
• Federated AI research across decentralized clients
• Privacy-preserving machine learning benchmarking
• Fairness and bias mitigation testing under distribution shift
• Integration with Cifer's federated orchestration engine and no-code workspace

---

📈 Performance

Trained on a synthetic dataset benchmarked against real-world financial logs. This model achieves 99.93% accuracy, closely matching the 99.98% benchmark of models trained on real financial data. Performance consistency is preserved across federated nodes when using FedAvg aggregation.

---

🔐 Privacy & Federated Context

• Designed for federated training across 4 dataset partitions
• No raw data sharing between clients or central servers
• Supports Cifer’s asynchronous training and client coordination
• Compatible with Cifer’s blockchain-based contribution tracking and aggregation module

---

🔧 File Info

• Format: .h5 (Keras model file)
• Input: Preprocessed numeric tabular data (StandardScaler + LabelEncoded type)
• Target: isFraud binary label (0 or 1)
• Recommended loader: keras.models.load_model("client_model.h5")

---

📜 License

Apache 2.0

---

🙌 Citation

If you use this model or dataset in your work, please cite:

CiferAI (2025). Cifer Fraud Detection Dataset & Federated Model – Privacy-Preserving AI for Financial Risk.

---