Data anonymization techniques involve removing or masking identifiable information from datasets to protect individual privacy [9]. Common methods include k-anonymity, l-diversity, and t-closeness, which attempt to make it difficult to trace data back to individuals. The advantages of data anonymization include simplicity and minimal computational demands, making it suitable for low-power IoT devices. However, limitations exist, as anonymized data can sometimes be re-identified through advanced techniques, especially when combined with other datasets, posing a risk in interconnected smart city systems.

Federated learning allows devices to train machine learning models locally on their data and then share only the model parameters with a central server rather than raw data [10]. This approach enhances privacy by keeping data decentralized. Federated learning is particularly useful in IoT environments where privacy is critical, enabling devices to contribute to data insights without compromising individual privacy. Advantages include reduced data transmission, preserving bandwidth, and minimizing data exposure. Limitations include higher device-side computational requirements and potential security risks, such as model inversion attacks.

Encryption protocols like End-to-End Encryption (E2EE) and homomorphic encryption provide security for IoT data by transforming it into an unreadable format accessible only to authorized parties [11]. Advantages include high data security and suitability for sensitive applications like health monitoring and public safety in smart cities. However, limitations are pronounced in IoT, where low-power devices often struggle with the computational load of strong encryption methods, and latency issues may impact real-time applications.

As smart cities grow, the number of IoT devices and the volume of data generated expand rapidly, requiring highly scalable privacy models. Many traditional privacy-preserving techniques struggle to accommodate such vast amounts of data, leading to latency issues or high computational costs that may be unsustainable for a large-scale IoT network.

Applications like traffic monitoring or emergency response demand real-time data processing, which is complicated by privacy-preserving protocols that can slow down data transmission and processing. Balancing the need for rapid data handling with privacy measures is challenging, especially for time-sensitive applications.

Smart city initiatives often promote transparency to foster trust with citizens, but the need for transparency can conflict with privacy-preserving practices. For instance, anonymization might limit transparency in data-sharing initiatives, creating a trade-off that city planners and administrators must carefully balance.

While encryption is a powerful tool for privacy, IoT devices in smart cities often have limited power and processing capabilities. This constraint requires lightweight, efficient encryption methods to secure data without overloading devices or increasing latency, particularly for energy-constrained settings.

The evolving nature of cyber threats necessitates adaptive models that can respond to new security challenges without requiring constant manual updates. Current approaches tend to be rigid, and research is needed to develop privacy models that can self-adjust based on contextual changes or emerging risks, ensuring ongoing privacy for sensitive data in smart cities.

The diversity of IoT devices and systems in smart cities makes implementing a unified privacy framework difficult. Gaps exist in creating standardized, interoperable privacy models that work seamlessly across various devices, operating systems, and vendors. Addressing this gap could enhance the efficiency of privacy-preserving techniques and streamline their implementation across smart city infrastructure.

This review highlights the strengths and limitations of existing privacy-preserving models. It identifies key areas where further research and development are essential to create robust, adaptable, and efficient solutions for the growing demands of smart city infrastructure.

The model uses E2EE to secure data from when it is collected by IoT devices until it reaches authorized endpoints. AES-256 encryption was selected for its balance between security strength and computational efficiency, which is suitable for the limited resources of IoT devices. Public-key cryptography is used for secure key exchange, ensuring only authorized entities can access encrypted data. This method safeguards data from interception or unauthorized access during transmission and storage.

Federated learning allows IoT devices to perform model training on local data and send only aggregated model updates to a central server rather than raw data. This approach minimizes central data storage, reducing exposure to potential breaches. TensorFlow Federated or similar frameworks support local device training, allowing for model training on a diverse set of edge devices without centralizing sensitive data. A federated averaging algorithm securely aggregates model updates, enhancing data privacy and model robustness.

To protect individual data even within aggregated outputs, differential privacy is applied by adding controlled noise to the model updates or aggregated data. Using Laplace or Gaussian mechanisms, this noise obfuscates specific data points to make re-identification difficult while preserving overall data utility. Adjustable privacy parameters balance model accuracy and privacy protection, allowing the approach to adapt based on data sensitivity.

The model will be implemented and tested in a simulated smart city environment, created using Python for data management and TensorFlow Federated for federated learning. AWS IoT Core or Google Cloud IoT simulates IoT device interactions, enabling controlled experimentation with a realistic setup. This environment simulates smart city data flows, from local data collection to centralized model aggregation, providing a foundation for testing privacy and performance.

Data encryption is implemented with the PyCryptodome library, which supports AES-256 and public-key cryptography. Encryption is applied before data transmission, and decryption occurs only when essential. Public-Key Infrastructure (PKI) manages keys, ensuring secure exchanges and storage across devices and servers.

TensorFlow Federated enables federated learning across simulated IoT devices, with each device processing its local data to produce model updates. A federated averaging algorithm aggregates model updates from multiple devices, creating a global model on the central server. Differential privacy is applied to model parameters, ensuring no sensitive information is revealed in the aggregated updates.

Synthetic data is generated to simulate real-world smart city applications such as traffic monitoring, environmental sensors, and utility management. Data types include GPS locations, environmental readings, energy consumption, and traffic patterns, reflecting typical smart city datasets. Sensitive attributes are flagged for special privacy processing, ensuring privacy is prioritized immediately.

Data preprocessing includes anonymizing sensitive fields and adding differential privacy noise to specific attributes, depending on the type and sensitivity of the data. Each IoT device performs preprocessing locally, reducing the chance of sensitive data exposure during transmission.

After preprocessing and encryption, data is transmitted from edge devices to the central server for model aggregation. Encryption ensures security during transit, while federated learning limits the need to share raw data. Aggregated model updates from federated learning are differentially private, adding another layer of security to protect against potential re-identification.

This methodology integrates privacy at each stage of the data lifecycle, leveraging encryption, federated learning, and differential privacy to safeguard data while supporting the functionality needed in smart city environments. The model is designed for scalability and adaptability, making it suitable for diverse, large-scale IoT implementations in real-world smart cities.

The proposed model demonstrated strong privacy preservation across the simulated smart city applications. E2EE ensured that sensitive data remained protected during transmission, with no recorded breaches or unauthorized access in transit. The federated learning approach prevented raw data from being centralized, significantly reducing privacy risks associated with data storage. Differential privacy effectively obscured individual data points within the aggregated model, ensuring that model updates resisted re-identification attacks. Privacy leakage tests showed that the model maintained a high privacy standard, with no discernible patterns that could be traced back to individual data sources.

In terms of efficiency, the model maintained a reasonable computational load on IoT devices. While E2EE did increase CPU usage on low-power devices by approximately 15%, the computational impact was within manageable limits. Federated learning reduced data transmission requirements by 35% compared to traditional centralized models, leading to significant bandwidth savings. Differential privacy settings, calibrated to balance privacy and data utility, had minimal impact on model accuracy, with only a slight 2% decrease in performance relative to a non-private model.

The latency impact from encryption and federated learning processes was within acceptable limits for most smart city applications, such as traffic monitoring and environmental sensing. Real-time applications (e.g., emergency response systems) experienced minor delays due to encryption and privacy computations, but the delay was consistently below 5ms, deemed acceptable for operational use.

The proposed model significantly improved privacy protection compared to traditional centralized models. Based on simulation tests, centralized models without federated learning and differential privacy showed a 28% higher risk of privacy breaches. The combined approach of federated learning and differential privacy in the proposed model reduced potential privacy leakage by over 30%, underscoring the effectiveness of multi-layered privacy mechanisms.

Thanks to federated learning, the model achieved a 35% reduction in bandwidth usage compared to centralized data storage approaches. Compared to traditional encryption-heavy models, the E2EE combined with federated learning allowed for quicker data aggregation while maintaining secure data handling, highlighting improved efficiency without sacrificing privacy.

Although federated learning adds complexity to deployment, the model's use of TensorFlow Federated and IoT emulation platforms streamlines implementation. Compared to existing approaches that rely solely on encryption or differential privacy, the integration of federated learning allows for faster data processing and privacy protection without needing extensive server storage infrastructure. The model's modular design makes it adaptable, facilitating easier adjustments in privacy parameters for different IoT devices and applications. Overall, the results highlight the proposed model’s robust privacy protection, efficiency gains, and practicality for IoT-based smart city applications. The multi-layered approach demonstrates meaningful improvements over existing models, offering a scalable, privacy-focused framework suitable for diverse smart city environments.