Optimizing the Privacy Accuracy Trade-off in Data Stream Mining Using Random Cycle Cumulative Noise Addition

Abstract

Data stream mining has become increasingly important as a result of the continuous generation of data in real world applications including healthcare monitoring, sensor networks, transportation systems and online services. While stream classification techniques provide strong predictive performance, publishing or processing such streams raises serious privacy concerns. Perturbation based privacy preserving data stream mining (PPDSM) methods address this issue by adding noise to the data before it is analyzed. Cumulative noise addition accumulates noise across a fixed-size cycle of records give a favourable privacy–accuracy trade-off compared with independent noise addition. However, a fixed cycle size introduces a predictable and periodic structure in the perturbation process. This paper proposes Random-Cycle Cumulative Noise Addition (RC-CNA), a perturbation mechanism that adds noise to random cycle sizes instead of fix-cycle size. RC-CNA is evaluated against the fixed-cycle model. The proposed method raises classification accuracy, lowers privacy breach probability, and achieves the overall Privacy–Accuracy trade-off. Proposed method indicates that randomizing the cycle size, combined with the modulation, is a practical improvement that simultaneously strengthens privacy protection and improves classification utility.

Introduction

The text presents Random-Cycle Cumulative Noise Addition (RC-CNA), a new privacy-preserving technique designed to protect sensitive information in continuous data streams while maintaining machine-learning classification accuracy.

Modern applications such as healthcare monitoring, financial transactions, and transportation services generate large amounts of data that must be processed in real time. However, these data streams may contain sensitive personal information, creating privacy risks when they are used for machine learning. Privacy-Preserving Data Mining (PPDM) addresses these risks by modifying data before analysis, often through noise addition. The main challenge is balancing privacy protection with classification accuracy, since excessive noise can reduce the usefulness of the data.

1. Existing Methods and Their Limitations

Traditional stream classification algorithms, such as Hoeffding Trees and Adaptive Random Forests (ARF), learn incrementally from incoming data and adapt to changing patterns. However, processing sensitive information continuously can expose individuals to privacy attacks.

Previous cumulative noise addition methods improve privacy by adding noise progressively to data records. However, unbounded noise can accumulate excessively and reduce classification accuracy. Later approaches introduced fixed-length noise cycles to control this accumulation, but their predictable cycle boundaries create vulnerabilities that attackers can exploit to reconstruct original data.

2. Proposed RC-CNA Method

The proposed RC-CNA technique addresses these limitations through three main mechanisms:

• Random cycle lengths: Instead of using fixed-length cycles, the system randomly selects each cycle length from a predefined range, making cycle boundaries more difficult for attackers to predict.

• Gradually decaying cumulative noise: Previously added noise fades over time rather than being abruptly reset to zero at each cycle boundary. This helps control long-term noise accumulation while avoiding sudden discontinuities.

• Adaptive noise scaling: The amount of noise is adjusted according to the local variability of the data. More variable data chunks receive a different noise scale, allowing the perturbation process to respond to changing data characteristics.

The system adds Gaussian noise to the non-target attributes of incoming records while preserving the target class labels for classification.

3. System Architecture and Evaluation

The data stream is processed in small chunks, with each chunk perturbed before being supplied to the classification model. The approach uses an Adaptive Random Forest classifier and prequential evaluation, in which each incoming record is first used to test the current model and then used to update it.

The privacy evaluation considers a known input–output attack, where an attacker has access to a limited number of original and perturbed record pairs and attempts to reconstruct unknown original records.

Two primary privacy metrics are used:

• Relative Reconstruction Error (RE): Measures the normalized difference between the original data and its reconstructed estimate. Higher values generally indicate greater resistance to reconstruction.

• Privacy-Breach Probability (PbreachPbreach?): Measures the proportion of attacked records reconstructed within a specified error threshold.

The combined Privacy–Accuracy Magnitude (PAM) score is defined as:

PAM=RE2+Pbreach2PAM=RE2+Pbreach2?

The text states that a lower PAM indicates a more favorable combined outcome under the study's metric, although the two terms should be interpreted carefully because higher reconstruction error generally improves privacy.

4. Research Contributions

The study makes three main contributions:

• Introduces random cycle lengths to reduce the predictable patterns exploited by known input–output attacks.

• Develops a noise mechanism that combines gradual noise decay, cycle-position modulation, and data-adaptive scaling.

• Proposes an empirical evaluation comparing RC-CNA with the fixed-cycle baseline, LRW, in terms of classification accuracy, reconstruction resistance, and PAM.

The authors report that RC-CNA is intended to improve privacy protection while preserving or improving classification performance. However, the provided text describes the methodology and claimed contributions rather than presenting the final experimental results needed to independently verify the improvements.

Conclusion

This paper proposed Random-Cycle Cumulative Noise Addition (RC-CNA), a privacy-preserving perturbation mechanism for data stream mining. It randomizes the cycle length used in cumulative noise addition instead of fix cycle size. RC-CNA was evaluated against the fixed-cycle using an Adaptive Random Forest classifier under prequential evaluation and a known input–output attack model. The proposed method improved classification accuracy on the perturbed stream, reduced relative reconstruction, reduced ?-privacy breach probability and reduced the overall Privacy–Accuracy Magnitude. These results confirm that randomizing cycle size is an effective and low overhead way to close the fixed-cycle vulnerability of cumulative noise addition while also improving classification utility.

References

[1] Domingos, P. and Hulten, G. 2000. Mining high-speed data streams. Proceedings of the Sixth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 71–80. [2] Gomes, H. M., Bifet, A., Read, J., Barddal, J. P., Enembreck, F., Pfharinger, B., Holmes, G. and Abdessalem, T. 2017. Adaptive random forests for evolving data stream classification. Machine Learning, 106(9-10), 1469–1495. [3] Denham, B., Pears, R. and Naeem, M. A. 2020. Enhancing random projection with independent and cumulative additive noise for privacy-preserving data stream mining. Expert Systems with Applications, 152. [4] Hewage, U. H. W. A., Pears, R. and Naeem, M. A. 2022. Optimizing the Trade-off Between Classification Accuracy and Data Privacy in the Area of Data Stream Mining. International Journal of Artificial Intelligence, 20(1), 147–167. [5] A. Cuzzocrea, Privacy-Preserving Big Data Stream Mining: Opportunities, Challenges, Directions, IEEE International Conference on Data Mining Workshops, 2017. [6] U. H. W. A. Hewage, R. Sinha, M. Asif Naeem, Privacy-preserving data (stream) mining techniques and their impact on data mining accuracy: a systematic literature review, Springer, 2023. [7] Liu, K., Kargupta, H. and Ryan, J. 2006. Random Projection-Based Multiplicative Data Perturbation for Privacy Preserving Distributed Data Mining. IEEE Transactions on Knowledge and Data Engineering, 18, 92–106. [8] Liu, K. 2007. Multiplicative Data Perturbation for Privacy Preserving Data Mining. PhD thesis, University of Maryland, Baltimore County.Giannella, C., Kargupta, H. and Liu, K. 2008. A Survey of Attack Techniques on Privacy-Preserving Data Perturbation Methods. [9] Giannella, C., Kargupta, H. and Liu, K. 2008. A Survey of Attack Techniques on Privacy-Preserving Data Perturbation Methods. [10] Feifei Li, Jimeng Sun, Hiding in the Crowd: Privacy Preservation on Evolving Streams through Correlation Tracking, IEEE, 2007. [11] Ching-Ming Chao, Po-Zung Chen, Chu-Hao Sun, Privacy-Preserving Classification of Data Streams, Journal of Science and Engineering, 2009. [12] M. A. P. Chamikara, P. Bertok, D. Liu, S. Camtepe, I. Khalil, Efficient data perturbation for privacy preserving and accurate data stream mining, Pervasive and Mobile Computing, 48 (2018) 1–19. [13] Fatlawi, Hayder K. and Kiss, Attila, Differential privacy based classification model for mining medical data stream using adaptive random forest, Acta Universitatis Sapientiae, Informatica, vol. 13, no. 1, 2021, pp. 1–20. [14] Waruni Hewage, R. Sinha, M. Asif Naeem, An accuracy-privacy optimization framework considering user\'s privacy requirements for data stream mining, 2025 [15] Jerzy Stefanowski, Dariusz Brzezinski, Stream Classification, in Encyclopedia of Machine Learning, Springer. [16] Albert Bifet, MOA data stream mining: A practical approach, vol 8, pp. 127-141, May 2011. [17] Bifet, A., Kirkby, R., Kranen, P. and Reutemann, P. 2012. Massive Online Analysis (MOA) Manual. [18] Kaggle. 2017. New York City Taxi Trip Duration. https://www.kaggle.com/c/nyc-taxi-trip-duration

Copyright

Copyright © 2026 Anita A. Parmar, Chirag A. Patel. 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.