An Ensemble Architecture Based on Deep Learning Model for Click Fraud Detection in Pay-Per-Click Advertisement Campaign

Abstract

Pay-per-click (PPC) advertising has become a foundational component of digital marketing, enabling advertisers to reach targeted audiences through keyword-based ad placements. However, the rise in automated scripts, bots, and click farms has led to a surge in click fraud—invalid clicks that artificially inflate advertising costs and mislead performance metrics. Existing fraud detection techniques often rely on static rule-based systems or shallow machine learning models, which are inadequate in identifying evolving and obfuscated fraudulent behavior. This paper proposes an ensemble deep learning architecture that integrates a Convolutional Neural Network (CNN) with a Long Short-Term Memory (LSTM) network for robust click fraud detection. The CNN module captures local temporal patterns within clickstream features, while the LSTM module models long-term behavioral dependencies across user sessions. Experimental results on benchmark datasets demonstrate that the proposed hybrid model outperforms conventional models in accuracy, precision, recall, and F1-score. The ensemble approach effectively reduces false positives while adapting to evolving fraud signatures, offering a scalable and intelligent solution for securing PPC platforms.

Introduction

The paper proposes a robust deep learning-based approach to detect click fraud in Pay-Per-Click (PPC) advertising systems using a hybrid CNN-LSTM ensemble model. Click fraud, which involves artificial inflation of ad clicks, is a significant challenge in digital advertising and often utilizes tactics like botnets and click farms. Traditional detection systems, relying on rule-based or shallow machine learning methods, struggle with adaptability, evolving fraudulent tactics, and high false positives.

Key Contributions:

• Hybrid deep learning model combining Convolutional Neural Networks (CNNs) for spatial feature extraction and Long Short-Term Memory (LSTM) networks for temporal sequence modeling.

• Feature engineering specific to clickstream data, such as click frequency, dwell time, and IP repetition.

• Comparative performance evaluation showing improved detection accuracy and fewer false alarms compared to traditional models.

• Modular framework suitable for real-time integration into PPC fraud monitoring systems.

Literature Survey:

• Traditional Approaches: Early methods were rule-based, focusing on anomalies like multiple clicks from the same IP. While computationally efficient, these systems lacked adaptability and were vulnerable to evasion techniques.

• Machine Learning Models: Supervised models like Decision Trees, SVM, and Random Forests improved detection but still faced issues with concept drift and reliance on manually engineered features.

• Deep Learning Models: CNNs and LSTMs have shown success in capturing complex, non-linear patterns and temporal dependencies in user behavior. However, single-model approaches often fail to generalize to diverse fraud scenarios.

• Ensemble Models: The paper highlights the potential of CNN-LSTM ensemble architectures, which combine the strengths of both CNN (spatial patterns) and LSTM (temporal patterns) to improve fraud detection.

Proposed Methodology:

The ensemble architecture includes:

1. Data Preprocessing: Includes missing value handling, feature encoding, timestamp conversion, and normalization.

2. Feature Engineering: Focus on capturing user behavior patterns like click frequency, dwell time, and IP repetition.

3. CNN Module: Extracts local patterns from structured input data (click behavior).

4. LSTM Module: Captures temporal dependencies across user sessions to detect long-term fraud behaviors.

5. Output Layer & Classification: Binary classification of fraudulent vs. legitimate clicks using a dense output layer with a sigmoid activation.

6. Training Strategy: The model is trained with an 80:20 train-test split, using metrics like accuracy, precision, recall, F1-score, and AUC-ROC for evaluation.

Implementation:

• Software Environment: Implemented in Python 3.10, using frameworks like TensorFlow/Keras for model construction and training, and Google Colab for GPU acceleration.

• CNN-LSTM Architecture: A hybrid model consisting of convolutional layers (for local pattern extraction) followed by LSTM layers (for temporal dependency modeling).

• Evaluation Metrics: The model's performance was evaluated using accuracy, precision, recall, F1-score, and AUC-ROC, with visualization tools to track performance (e.g., ROC curves, confusion matrix).

Results:

• The CNN-LSTM ensemble model demonstrated superior classification performance in detecting click fraud compared to conventional models. It achieved high accuracy and reduced false alarms, providing a more reliable solution for PPC fraud detection in dynamic, real-world conditions.

Conclusion

Click fraud continues to undermine the trustworthiness and economic efficiency of Pay-Per-Click (PPC) advertising platforms, inflicting substantial financial losses on advertisers and distorting campaign analytics. To address this persistent challenge, this study proposes a deep learning-based ensemble architecture that integrates Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks for accurate and adaptive fraud detection. The CNN-LSTM model combines local pattern recognition with temporal behavioral analysis, enabling the detection of complex, evolving click fraud strategies that static models often fail to capture. Experimental evaluation demonstrated that the proposed model outperforms traditional classifiers and standalone deep learning models in terms of accuracy, precision, recall, F1-score, and AUC-ROC. The ensemble framework also maintains a low false positive rate, making it suitable for deployment in real-time advertising environments where both sensitivity and specificity are critical. Beyond technical performance, the proposed system is scalable, modular, and easily integrable into existing ad monitoring pipelines. Its capacity to learn from clickstream dynamics without requiring manual feature engineering ensures resilience against adversarial manipulation and concept drift. In conclusion, the CNN-LSTM ensemble architecture provides a robust, intelligent solution for detecting click fraud in PPC campaigns, thereby enhancing advertiser protection, improving ad budget efficiency, and strengthening the integrity of digital marketing ecosystems.

References

[1] Z. Liu, Y. Li, and H. Liu, “Click fraud detection on the advertiser side: A deep learning approach,” in Proc. 2019 Int. Conf. Big Data (Big Data), Los Angeles, CA, USA, 2019, pp. 2959–2968. [2] J. Hu, Y. Wang, and X. Jiang, “Sequential behavior modeling for click fraud detection using LSTM networks,” in Proc. 2020 Int. Joint Conf. Neural Networks (IJCNN), Glasgow, UK, 2020, pp. 1–8. [3] B. Sun, T. Wang, and K. Qin, “Ensemble learning for fraud detection using hybrid deep networks,” IEEE Access, vol. 9, pp. 109387–109399, 2021. [4] S. Kumar and M. Patel, “Detecting click fraud in pay-per-click advertising using behavioral analysis,” in Proc. 2021 Int. Conf. Intelligent Computing and Control Systems (ICICCS), Madurai, India, 2021, pp. 365–370. [5] Y. Kim, H. Lim, and J. Kang, “Ad fraud detection with deep learning: A survey,” IEEE Access, vol. 10, pp. 18991–19006, 2022. [6] S. Hochreiter and J. Schmidhuber, “Long Short-Term Memory,” Neural Computation, vol. 9, no. 8, pp. 1735–1780, 1997. [7] K. Simonyan and A. Zisserman, “Very Deep Convolutional Networks for Large-Scale Image Recognition,” arXiv preprint, arXiv: 1409.1556, 2014.

Copyright

Copyright © 2025 Ch. Divya, Nalamalapu Gayathri, Gutti Bhagyalakshmi, Devathoti Pavithra, Metla Devi Aditya Prasad, Nallapaneni Manil. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.