Generative Artificial Intelligence for Internet of Things Computing: A Systematic Survey

As an excellent, fastidious, and diligent researcher, I have meticulously analyzed both provided texts concerning the survey paper on Generative Artificial Intelligence for Internet of Things Computing. My task is to synthesize these findings into a long, detailed summary that accurately reflects the scope, methodology, key findings across different perspectives (network-, object-, and semantic-oriented), challenges, and future directions outlined in the original paper.

Here is the comprehensive synthesis:

Detailed Research Synthesis: Generative Artificial Intelligence for Internet of Things Computing: A Systematic Survey

This systematic literature review addresses a critical gap in current research by providing a holistic overview of the opportunities, inherent issues, and considerations arising from the convergence of Generative Artificial Intelligence (GenAI) and the Internet of Things (IoT) computing. The authors explicitly state their aim is to provide this comprehensive view by offering a systematic survey following the rigorous PRISMA methodology. They confirm that an initial informal search, supplemented by personal experience, validated a substantial body of existing contributions, underscoring the necessity for such a systematic review to map these diverse applications and challenges.

The core contribution of this work is its structured comparison framework and the definition of well-defined research questions designed to explore the past, present, and future trajectories of GenAI integration with IoT Computing. The survey spans 74 eligible studies published between 2020 and 2025, ensuring a contemporary view of this rapidly evolving intersection.

Categorization and Scope (RQ1)

The review is systematically organized into three distinct domains—network-oriented, object-oriented, and semantic-oriented—to effectively highlight the unique features and overlapping aspects of GenAI's application across these dimensions. This structure directly addresses Research Question 1 (RQ1): What are the current scopes of GenAI for IoT applications?

The findings reveal a broad spectrum of applications:

• Networking Domain: GenAI is actively employed in specific networking tasks, including the generation of synthetic test data for 5G networks [105], mobile edge networks [61, 127], and node prediction within blockchain networks [110]. Furthermore, research explores concepts like GAI-IoV (Bridging Generative AI and Vehicular Networks for Ubiquitous Edge Intelligence) [118] and the power of edge-cloud generative AI in mobile networks via a survey of AIGC services [120].

• Object Domain: Applications are evident in areas such as smart home systems [90], urban environmental learning [102], and traceability/performance optimization within processes like the recycling of WEEE, where GenAI, Digital Twins, and DRL are utilized [111].

• Semantic Domain: A significant portion of the analyzed studies utilizes implicit density models, particularly Generative Adversarial Networks (GANs), within IoT contexts.

Key Models and Architectures (RQ2)

The survey comprehensively addresses Research Question 2 (RQ2): Which are the main GenAI models and architectures that support IoT Computing? The analysis categorizes Density Models into two primary forms:

1. Explicit Density Models: These include tractable approaches such as Auto Regressive Models (ARMs), Diffusion Models, Variational Auto-Encoders (VAEs), and Energy-based Models [49, 56, 60, 97].

2. Implicit Density Models: A significant number of studies leverage models like Generative Adversarial Networks (GANs) within IoT applications.

The paper emphasizes that Discriminative AI models, while instrumental in supporting the design and implementation of intelligent IoT applications by extracting patterns from large datasets for probabilistic classification, are being moved beyond their limitations. The authors advocate for a broader perspective where probability estimation serves as a foundation to simultaneously handle uncertainty and understanding, which is precisely what GenAI offers through its ability to generate context-aware and adaptable content based on learned representations.

Challenges and Research Gaps (RQ3)

The most frequently reported challenge across the analyzed studies pertains to resource constraints. This is a critical barrier to the full adoption of these advanced models in real-world IoT scenarios. Specific constraints highlighted include:

• Limited computational power, memory availability, and energy constraints within IoT and edge environments.

• Security risks associated with deploying GenAI within IoT systems [119].

The survey clearly identifies a significant gap: comprehensive and domain-independent secondary studies on the GenAI-IoT duo are very few. This lack of comprehensive literature necessitates the current review.

Future Research Directions (RQ4)

Addressing Research Question 4 (RQ4): Which are the future research directions? The authors strongly conclude that as IoT grows, GenAI will play a crucial role in scaling and managing its complexity. To realize this potential, several key optimization and adaptation techniques have been highlighted:

1. Model Optimization: Improving training efficiency, reducing computational costs, and enhancing the adaptability of GenAI models for resource-constrained systems. This includes model compression techniques such as Pruning, LoRA, Quantization, Knowledge Distillation (KD), and Automated Design.

2. On-Device Training Techniques: Exploring Parameter-Efficient Fine-Tuning (PEFT), Memory-Efficient Fine-Tuning (MEFT), and Data-Efficient Fine-Tuning (DEFT).

3. Offloading Strategies: Implementing efficient workload partitioning and optimized communication protocols to manage computational load across distributed systems.

4. Distributed Training and Privacy: Federated Learning (FL) is presented as a vital method for addressing both privacy preservation and enabling scalable, distributed model training across IoT devices.

Synthesis of Optimization Techniques

The paper meticulously details the methods proposed to mitigate resource constraints:

• Model Compression: Techniques like Pruning, LoRA, Quantization, KD, and Automated Design are discussed extensively to reduce the computational complexity of Density Models (DGMs).

• Efficient Fine-Tuning: PEFT, MEFT, and DEFT are explored for training models directly on devices.

• Distributed Learning: Federated Learning (FL) is emphasized for its dual benefit: enhancing scalability while preserving data privacy during model training.

Conclusion

In summary, the survey confirms that GenAI is not merely a theoretical concept but is actively being deployed across diverse IoT facets—from networking and edge intelligence to environmental learning and process optimization. While the literature demonstrates significant application potential, the central theme remains the deployment challenge imposed by resource limitations. The paper successfully maps these findings using structured frameworks (Table 10 for RQ mapping, Table 8 for Object-oriented comparisons, Table 7 for Semantic-oriented comparisons), providing a robust roadmap detailing current models (explicit and implicit DGMs), prevalent challenges (resource constraints and security), and actionable future research avenues focused heavily on model compression, efficient fine-tuning, and distributed learning strategies. The authors firmly conclude that GenAI's role will be pivotal in managing the escalating complexity inherent in growing IoT ecosystems.

1. Model compression techniques can be applied to Deep Generative Models (DGMs) to reduce their computational complexity and energy consumption, as suggested by Pruning, Low Rank Approximation (LoRA), and Quantization. This allows DGMs to run on resource-constrained IoT devices while maintaining performance.

2. Knowledge distillation can be used to transfer knowledge from a complex teacher model to a more compact student model, which is particularly useful for deploying models on edge devices where computational resources are limited. This preserves essential generative features such as style or content consistency in smaller models.

3. Data-free quantization can be employed to reduce bit precision without requiring access to training data, offering a solution for privacy-constrained applications in IoT domains where training data access is restricted. This technique allows for model compression while maintaining accuracy without needing the full training dataset.

4. Parameter-Efficient Fine-Tuning (PEFT) techniques, such as the Addition-Based Approach or Reparameterization Based Approach, can be used to reduce the computational cost of fine-tuning DGMs on devices by selecting only a subset of essential parameters for tuning. This allows for adaptation to new unseen data in dynamic IoT environments while minimizing the runtime memory footprint.

5. Federated Learning (FL) with Generative AI techniques can enable privacy-preserving and distributed training across multiple edge devices, addressing the challenge of training massive DGMs without centralizing sensitive raw data. This synergy ensures that sensitive information remains local on IoT devices while allowing for collaborative model development.

6. Offloading strategies, such as work load partitioning and efficient communication protocols, can be used to distribute the execution of entire DGMs between IoT devices and edge servers or cloud infrastructure. This enables the use of large-model DGMs by balancing load across heterogeneous resources.

The integration of Generative Artificial Intelligence (GenAI) within the Internet of Things (IoT) is garnering considerable interest. This growing attention stems from the continuous evolution and widespread adoption they are both having individually, enough to spontaneously reshape numerous sectors, including Healthcare, Manufacturing, and Smart Cities. Hence, their increasing popularity has catalyzed further extensive research for understanding the potential of the duo GenAI-IoT, how they interplay, and to which extent their synergy can innovate the state-of-the-art in their individual scenarios. However, despite the increasing prominence of GenAI for IoT Computing, much of the existing research remains focused on specific, narrowly scoped applications. This fragmented approach highlights the need for a more comprehensive analysis of the potential, challenges, and implications of GenAI integration within the broader IoT ecosystem. This survey exactly aims to address this gap by providing a holistic overview of the opportunities, issues, and considerations arising from the convergence of these mainstream paradigms. Our contribution is realized through a systematic literature review following the PRISMA methodology. A comparison framework is presented, and well-defined research questions are outlined to comprehensively explore the past, present, and future directions of GenAI integration with IoT Computing, offering valuable insights for both experts and newcomers.