Privacy-Preserving Machine Learning for IoT: A Cross-Paradigm Survey and Future Roadmap
Organizations: School of CSE, University of NSW, AU · TCS Research & Innovation, Tata Consultancy Services (Australia & New Zealand), Australia
Abstract
The rapid proliferation of the Internet of Things has intensified demand for robust privacy-preserving machine learning mechanisms to safeguard sensitive data generated by large-scale, heterogeneous, and resource-constrained devices. Unlike centralized environments, IoT ecosystems are inherently decentralized, bandwidth-limited, and latency-sensitive, exposing privacy risks across sensing, communication, and distributed training pipelines. These characteristics render conventional anonymization and centralized protection strategies insufficient for practical deployments. This survey presents a comprehensive IoT-centric, cross-paradigm analysis of privacy-preserving machine learning. We introduce a structured taxonomy spanning perturbation-based mechanisms such as differential privacy, distributed paradigms such as federated learning, cryptographic approaches including homomorphic encryption and secure multiparty computation, and generative synthesis techniques based on generative adversarial networks. For each paradigm, we examine formal privacy guarantees, computational and communication complexity, scalability under heterogeneous device participation, and resilience against threats including membership inference, model inversion, gradient leakage, and adversarial manipulation. We further analyze deployment constraints in wireless IoT environments, highlighting trade-offs between privacy, communication overhead, model convergence, and system efficiency within next-generation mobile architectures. We also consolidate evaluation methodologies, summarize representative datasets and open-source frameworks, and identify open challenges including hybrid privacy integration, energy-aware learning, privacy-preserving large language models, and quantum-resilient machine learning.
Figures & tables
| Acronym | Expansion |
|---|---|
| AUC | Area Under the Curve |
| BF | Bloom Filter |
| CDP | Central Differential Privacy |
| DL | Deep Learning |
| DML | Distributed Machine Learning |
| DNN | Deep Neural Network |
| Reference | Focus/Scope | Main Contributions | Limitations / Gaps |
|---|---|---|---|
| Amiri-Zarandi et al. , 2020 ( Amiri-Zarandi et al., 2020 ) | ML -based privacy in IoT (general) | Early survey linking ML techniques with IoT privacy; broad overview of solutions. | Does not systematise formal PPML paradigms ( DP , FL , HE , synthetic); lacks attack–defence, datasets, and frameworks. |
| Al-Garadi et al. , 2020 ( Al-Garadi et al., 2020 ) | ML / DL ( DL ) for IoT security | Extensive review of ML / DL in IoT security context. | Security-centric; privacy-preserving aspects only briefly covered. |
| Jiang et al. , 2021 ( Jiang et al., 2021 ) | Differential Privacy in Industrial IoT (IIoT) | Detailed review of DP techniques for IIoT . | Domain-specific ( IIoT ); excludes broader IoT contexts, other paradigms, and datasets. |
| Abbas et al. , 2022 ( Abbas et al., 2022 ) | Safety, security, privacy in ML -based IoT | General overview of safety, security, and privacy concerns in ML -enabled IoT . | Not a formal PPML survey; lacks taxonomy, evaluation metrics, or datasets/frameworks. |
| Aouedi et al. , 2024 ( Aouedi et al., 2025 ) | Intelligent IoT : applications, security, privacy | Comprehensive overview of IoT applications including security/privacy issues. | Very broad scope; PPML methods treated only at high level. |
| Le et al. , 2024 ( Le et al., 2025 ) | Distributed ML in IoT | Covers architectures and applications of distributed ML in IoT . | Distributed ML focus; privacy-preserving aspects underexplored. |
| Paradigm | Privacy | Utility | Robustness |
|---|---|---|---|
| Perturbation-Based ( DP ) | , ; MIA AUC | Accuracy; RMSE ; MAE | FGSM ASR; perturbation norm |
| Distributed Learning ( FL ) | Per-round ; cumulative DP budget; MIA ASR | Convergence rate; accuracy under non-IID | Byzantine resilience; gradient norm |
| Cryptographic ( HE / SMPC ) | Formal security parameter; leakage probability | Latency; throughput; RMSE | Side-channel resistance |
| Generative ( GAN ) | MIA AUC ; reconstruction risk | KL ; MMD ; DTW ; TSTR | Mode collapse rate; FGSM ASR |
| Framework | Backend | DP Support | FL Support | Privacy Accounting | IoT Suitability |
| TensorFlow Privacy ( Abadi et al., 2016 ) | TensorFlow | DP-SGD , Gaussian, Laplace | Via TFF | Rényi DP ; | Moderate — TF Lite compatible |
| Opacus ( Yousefpour et al., 2021 ) | PyTorch | DP-SGD ; per-sample gradients | Limited | Rényi DP ; PRV accountant | Moderate — edge PyTorch deployments |
| Flower (FLwr) ( Beutel et al., 2020 ) | Agnostic (TF, PyTorch, others) | Via integration with Opacus/TFP | Native; heterogeneous devices | Via external accountants | High — designed for edge/embedded |
| PySyft ( Ziller et al., 2021 ) | PyTorch | DP via integration | Native SMPC ; FL | Limited built-in | Low — high compute requirements |
| FATE ( Liu et al., 2021 ) | Standalone | DP via integration | Native; cross-silo FL | Limited built-in | Low — enterprise-oriented |
Privacy-Preserving ML
Privacy-preserving machine learning develops methods to learn, share, and run models while limiting exposure of sensitive data.
See all 433 papers