Content Centric Routing

Content-centric routing, a core feature of ICN protocols, revolutionizes network architecture by shifting the responsibility of name resolution from the application layer to the network layer [34]. This innovation proves particularly advantageous in fragmented networks, where traditional location-based addressing may falter due to irregular or intermittent connectivity disruptions. For example, in environments without fixed network infrastructure, where connections are ad hoc, ICN's content name-based routing allows content to be replicated based on its name rather than its physical location. This means content can be efficiently distributed without needing servers like DNS resolvers at the application layer.

Content-Based Access Management

ICN empowers controlling data access, restricting it to specific users or groups through content-based security measures [35]. This capability enhances secure exchanges among peer users, particularly in remote network regions. For instance, consider a financial institution utilizing ICN. With content-based access management, sensitive financial data can be restricted to authorized personnel only. This ensures that confidential information remains secure and accessible only to those with proper clearance, even in remote branches with limited connectivity.

Data Objects Integrity

ICN architecture revolves around named data objects, where various proposals within ICN incorporate the concept of ‘self-certifying data’ into their naming schemes [36]. This involves embedding cryptographic information about the data's origin, enabling authentication without external entities. For instance, in a distributed sensor network utilizing ICN, each sensor could embed cryptographic signatures in its data packets, allowing recipients to verify the authenticity of the data without relying on centralized authentication servers. Consequently, security and reliability are significantly enhanced within ICN architectures.

In-Network Caching

In-network caching, a fundamental aspect of ICN, involves storing frequently accessed content strategically within the network infrastructure [37]. This approach effectively handles large volumes of traffic, thereby reducing network congestion. For example, caching at access nodes can mitigate congestion in backhaul links by delivering content from nearby caches.

Dynamic Networking

In modern networking, ICN eliminates the need for a continuous end-to-end connection [38]. The characteristic of ICN eliminating the need for a continuous end-to-end connection promotes seamless integration of conventional network setups and fragmented alternatives. ICN enhances data retrieval and distribution by reducing dependency on constant direct connections in traditional, stable network environments, such as home Wi-Fi or LANs. Meanwhile, in dynamic and less stable networks, like mobile or ad-hoc networks in disaster recovery scenarios, ICN allows devices to intermittently connect without disrupting communication. Data can be cached and retrieved from multiple points, ensuring continuous access even when network paths are unstable. This versatility makes ICN a robust solution for various networking challenges, enhancing efficiency and reliability across different network types.

Naming

In the realm of ICN, a significant derivation from conventional networking methodologies is observed, particularly in how data is identified and accessed. Unlike traditional networks relying on hierarchical URLs, ICN uniquely assigns identities to data based on their names rather than their physical locations [41]. This shift enables a more versatile and dynamic approach to data retrieval, wherein flat names are frequently utilized to represent data. However, in the earlier ICN architecture, such as CCN, a distinct hierarchical organization of names prevails, enhancing readability and accessibility for users. This characteristic underscores ICN's flexibility, accommodating both flat and hierarchical naming conventions to suit diverse network architectures and user preferences.

Caching

Caching is a notable feature in ICN networks, facilitating content storage directly within network nodes. Each content item has the potential to be cached along the delivery route within these network nodes. Consequently, when subsequent requests for the same content arise, it can be efficiently delivered through the caches of these ICN nodes [42]. Implementing in-network caching occurs within ICN nodes between the content consumer and the content producer. This strategic approach not only reduces the workload on the content producer but also enhances data availability. Furthermore, in-network caching is pivotal in balancing load within the network, thereby decreasing content retrieval latency.

Routing

Routing in the ICN framework represents a strategic approach to managing numerous requests for the same content originating from a diverse user set. The process adheres to a specific protocol, wherein only the initial request is transmitted to a potential content source, and subsequent requests are systematically recorded and temporarily stored within the memory of the ICN node [43]. When the data is received from the source, it is then disseminated to each requester through the same interface where their original requests were initiated. As a result, ICN can inherently efficiently execute multicast, optimizing content delivery to multiple users.

Mobility

In the domain of ICN, mobility remains unaffected by the absence of a fixed connection structure. ICN operates on a model where no established connections exist, meaning that users can move freely without disrupting their access to content [44]. ICN's request-driven approach treats each request independently, allowing seamless content retrieval for mobile users, regardless of their location or when the request is made. This design, crafted for mobility, ensures continuous and efficient data access in dynamic environments.

Security

ICN provides robust self-protected security by utilizing encrypted and self-certified content, ensuring that only authorized users can access the content [45]. Unlike traditional methods where the connection between the data producer and the consumer is secured, ICN employs a distinct security approach. In ICN, the content producer incorporates a signature directly into the content. Consequently, consumers and intermediate caching points verify the legitimacy of the content using specific keys published by the content producer.

Hierarchical Naming

Hierarchical naming involves the construction of a name by assembling various components to represent an application's function, resembling the structure found in web addresses [46]. This approach simplifies the creation of easily comprehensible names and supports the organized management of many names. Nevertheless, challenges arise when addressing very small entities.

Flat Naming

In contrast to hierarchical names, flat names lack inherent meaning or structure. Their unconventional nature makes them less intuitive for people to comprehend, as they don't resemble typical words [47]. These names present difficulties when naming frequently changing entities, as they are not well-suited for new or dynamic content.

Attribute-Value-Based Naming

A name based on attributes and values is similar to having a collection of specific details about something [48]. Each detail is associated with a name, a category indicating the type of detail, and a list of potential options (such as creation date, content type, year, etc.). When these details are combined, they form a distinct name for the item and provide additional information about it. This naming system proves beneficial when searching for items using specific keywords. However, ensuring a quick and precise match isn't always straightforward due to the possibility of multiple things sharing the same keywords.

Hybrid Naming

A hybrid name involves blending various naming approaches discussed earlier to optimize the network's performance, enhance speed, and ensure information security and privacy [49]. For example, incorporating features from one method can facilitate efficient searches by intelligently combining names; utilizing concise and straightforward names from another method can conserve space, and adding details aids in keyword-based searches while maintaining security. However, crafting these unique hybrid names can be challenging, particularly when dealing with dynamic entities like real-time information.

Content Centric Networking (CCN)

CCN is acknowledged as one of the most popular and user-friendly architectures in Information Centric Networking [50]. In CCN, a client initiates an information request within the network by transmitting an ‘INTEREST,’ which travels through a series of routers or nodes. These nodes communicate with their neighboring nodes to locate the requested information. If the desired content is found at any of these nodes, it is then directed back to the requester. Importantly, in CCN's operation, all intermediate forwarding nodes store a copy of the requested data in their memory for potential future requests along the same path. In cases where the content is not discovered along the path, caching occurs at the central server.

Data Oriented Networking Architecture (DONA)

The DONA architecture is recognized as a notable platform within the domain of ICN, choosing flat names over hierarchical naming. This architectural choice presents numerous advantages, primarily focusing on the capability to cache information at the network layer. This is made possible through innovative concepts such as autonomous systems and LCD integrated into its design. The use of AS ensures the seamless retrieval of information, even in scenarios where nodes are temporarily unavailable due to mobility, as the information is cached within a node incorporated into the AS.

DONA's robust cryptographic capabilities empower users to authenticate their requests while it still relies on IP and routing addresses at both local and global levels [51]. A pivotal component in DONA is the unique server known as RH, which plays a critical role in resolving names as network operations progress. Integrated into the AS, RHs and routers operate under the network's routing policies.

In DONA, like CCN, data is registered and cached by RHs within the network. When a publisher node wants to make data available, it sends a “REGISTER” message alongside the data to an RH. The RH then stores a copy of this data locally and keeps track of its location. This caching process allows RHs to quickly provide the data to other RHs within the network hierarchy without needing to fetch it again from the original source. By notifying the entire RH hierarchy about the data's availability, DONA ensures efficient data distribution and retrieval across its network segments.

Publish-Subscribe Internet Routing Paradigm (PSIRP)

In PSIRP, when subscribers want to access information objects, they send a request to a rendezvous handler. This handler acts as a mediator, facilitating the process of finding and retrieving the requested data [52]. The request travels through the nearest router, which checks if the data is available either in its cache or directly from the publisher node.

At the same time, the publisher node registers the data with its own rendezvous handler. This handler then collaborates with the subscriber's handler to locate a match [53, 54]. Once a match is found, the data is routed through internal routers along the registered path to fulfill the subscriber's request.

Network of Information (NetInf)

The NetInf architecture distinguishes itself within the ICN landscape by utilizing a publisher/subscriber model facilitated by the NRS [55]. Subscribers access data by performing ‘look-ups,’ where publishers NDO with specific key attributes. To efficiently manage data transmission, NetInf utilizes the multicast DHT routing algorithm. This approach ensures that consumer nodes consistently receive updated lists of requests. The NRS stations are crucial in optimizing the delivery of NDO, ensuring efficient and reliable information retrieval across the network. The NRS stations play a pivotal role in optimizing the delivery of NDO.

Name Data Networking (NDN)

NDN comprises three fundamental components: producers, consumers, and NDN routers. Producers function as content distributors within the NDN framework. At the same time, consumers act as content subscribers, interacting with NDN elements such as CS, PIT, FIB, Faces, and the Forwarding Daemon to fulfill user requests for desired content [56, 57], as specified in Tables 1.

In the NDN architecture, communication occurs through INTEREST and DATA packets, as depicted in Figure 3.

CacheFilter

To mitigate data redundancy and support extensive content distribution in the network, the CacheFilter cache placement scheme is introduced in [64]. This scheme incorporates an additional field within the DATA packet known as the flag bit. A flag bit of 0 indicates that the content is already cached by a router within the network, while a flag bit of 1 signifies that the content needs to be cached.

Random Caching (RC)

RC is a method wherein nodes selectively cache incoming packets according to a predetermined static probability denoted as “p,” where 0 $$$ < $$$ p $$$ <= $$$ 1 as discussed in [65]. The commonly selected value for p is often established at 0.5. This widely adopted value seeks to reduce data redundancy, achieving a balance that mitigates both excessive duplication and underutilization of available resources.

Prob(p)

The objective of Prob (p) is to decrease caching redundancy and improve efficiency. In Prob(p), each content router is programmed to cache a data packet with a certain probability, denoted as p in work [66]. Upon the arrival of a data packet at a content router, a random number within the range of zero to one is generated. If the generated random number is less than p, the content router caches a copy of the data packet. Otherwise, the data packet is forwarded without being cached.

Probcache and Probcache+

ProbCache incorporates two components in the packet headers: one in the Interest-packet header and the other in the Data-packet header. The ratio between them is denoted as the CacheWeight for specific content. Probcache computes both cache weight and TimesIn, which refers to the sum of time a packet is saved on the caching path, subsequently determining the caching probability at each content router using a defined equation as presented in [67]. A content router with a higher caching probability than others along the path will cache contents. ProbCache+ represents an enhanced version of ProbCache and, similar to ProbCache, calculates TimesIn and CacheWeight at each content router. A content router with a higher caching probability than others along the path will cache contents with greater likelihood.

Leave Copy Down (LCD)

The purpose of LCD, as described in [68], is to reduce caching redundancy by introducing a caching proposal flag bit in the header of the data packet. The activation of the flag bit occurs when either the content source or a content router, having a replica of the corresponding content, receives a request. Utilizing this bit, a copy of the intended content is retained at the content router situated one level closer to the requester following each request.

Move Copy Down (MCD)

MCD, as discussed in [69], is employed to reduce caching redundancy, similar to LCD. In MCD, a duplicate of the content is transferred to the content router one level closer to the requester following each request.

Chunk Caching Location and Searching Scheme (CLS)

CLS, presented as an enhanced version of MCD in [70], aims to improve server workload efficiency. Consequently, CLS adopts a strategy where only a single replica of a chunk is maintained along the pathway connecting a server and a requester. Upon receiving a request, the chunk's copy moves one level closer to the requester. Conversely, during cache eviction, the chunk moves one level towards the server.

Chunk-Based Caching Scheme (WAVE)

The WAVE in the CNN framework is introduced by [71]. In this placement scheme, data packets are divided into uniform, fixed-sized fragments or chunks, which are then distributed across the network. WAVE's strategy combines the evaluation of content popularity and inter-chunk relationships as key factors in determining which content to cache. It strategically decides which specific chunks to cache based on the popularity of the content and the relationships between different chunks within the same content.

Most Popular Content (MPC)

A caching scheme called MPC was presented in [72]. This novel method entails edge devices actively monitoring and recording the frequency of requests for specific content names. These devices maintain a Popularity Table, essentially serving as a local repository that records the popularity metric for each content item. As the request counts accumulate, content names surpassing a fixed Popularity Threshold are recognized as popular. Once a content name attains this status, the hosting node for that content triggers a suggestion mechanism. This mechanism entails recommending neighboring nodes to cache the recognized popular content.

Dynamic Fine-Grained Popularity-Based Caching (D-FGPC)

Reference [73] has introduced an extended version of MPC named D-FGPC. D-FGPC incorporates a flexible popularity threshold function, determined by both the frequency of received INTEREST and the available cache capacity. Integrating this dynamic threshold function enhances the adaptability of the caching strategy employed by D-FGPC.

Popularity-Aware Closeness Caching (PaCC)

The goal of PaCC, as introduced by [74], is to systematically manage caching decisions by using the ‘PaCC’ field in the DATA packet. This aims to ensure efficient forwarding of content by identifying and utilizing the optimal router or node based on factors such as popularity and network characteristics, including hop distance.

Dynamic Popularity Cache Placement DPCP Scheme (DPCP)

The DPCP, as proposed in [75], is designed to manage the caching of data packets at a node efficiently. Upon the arrival of a data packet, a popularity calculation mechanism is triggered. This calculation incorporates the content's historical popularity in the last cycle and factors in the total number of requests in the current cycle. The DPCP scheme dynamically adjusts the threshold size through the AIMR algorithm, considering the current space occupation. Only content surpassing the current popularity threshold is cached, and the system strategically prioritizes the caching of content with a high probability of future requests.

Caching Popular and Fresh Content (CPFC)

The proposed CPFC caching decision scheme is introduced by [74]. It is a systematic approach for (CR) to manage the content cache effectively. Two significant metrics are employed by each CR: the Popularity metric, consisting of content prefixes and their corresponding request counts, and the Freshness metric, including content names and their respective lifetimes (times during which the content remains valid).

Two Layers Cooperative Caching (TLCC)

In the context of TLCC, as described in [76], CCN routers are grouped into Upper-Layer Group (ULG) and Lower-Layer Group (LLG) within an Autonomous System. CDULG caches received chunks based on hash comparison with the router's key range, while CDLLG caches chunks if the Local Popularity Count exceeds a threshold and the hash matches the router's key range.

DeepCache

Introducing the DeepCache framework in [77], highlighting the efficacy of LSTM-based models in predicting content popularity. The research addresses content popularity prediction as a seq2seq modeling problem and introduces the LSTM Encoder-Decoder model, a novel approach. Additionally, the DeepCache framework is presented for end-to-end cache decision-making.

CRCache

Reference [78] introduced a caching placement scheme called CRCache. This approach involves making caching decisions by considering a comprehensive popularity metric determined by assessing the frequency of received INTEREST. This metric is thoughtfully combined with identifying popular routers within the network topology. Notably, routers with superior distribution are marked as popular routers in the CRCache scheme.

Edge Caching

Edge caching strategically places content closer to users by storing it in the latest content router along a distribution pathway, as introduced by [79]. The goal is to minimize the number of hops required to access a content source, resulting in reduced time for content delivery and decreased network traffic caused by the transmission of content requests.

BEACON

To optimize network resources, a betweenness-centrality-based content placement scheme called BEACON, introduced by [80], employs a time series Gray model to predict which content to cache and where to cache it.

Content Betweenness Centrality (CBC)

The objectives of CBC, as detailed in [81], aim to minimize the latency associated with content delivery, mitigate traffic and congestion through an increased cache hit rate, and simultaneously reduce server load. CBC indicates the betweenness centrality exhibited by each content router within the network. In this context, data packets are strategically stored at the content router deemed ‘important,’ showing the highest betweenness centrality rate along the delivery path.

Content Selection and Router Selection (CaRS)

Reference [82] proposes a CaRS caching strategy for network devices aimed at balanced content distribution. It selects cached contents based on Zipf's law, and to balance content distribution, routers monitor neighboring cache status using parameters like proportionate distance from the consumer, router congestion, and cache status.

Link Congestion and Lifetime Based In-Networking Caching Scheme (LCLCS)

The main purpose of LCLCS is to maintain an acceptable cache rate and notably optimize network delay. The LCLCS process, defined in [83], begins with a network analysis to identify ‘important’ content routers along the content delivery path. Subsequently, LCLCS utilizes the computed lifetime to make decisions regarding the eviction of currently cached content, ensuring adequate space for caching incoming content.

Deep Learning-Based Content Popularity Prediction (DLCPP)

DLCPP, detailed in [84], proposed an auto-encoder-based popularity model implemented over the SDN controller to predict content popularity. It collected the data from the routers having spatial-temporal correlation to represent the flow of requested content at discrete intervals. Finally, a softmax classifier is employed to predict the popularity of content. Further, for popular placement, SDN selects a set of important routers with the help of the betweenness centrality approach.

Controller-Aware Secured Content Caching (CaCC)

CaCC, presented in [85], utilizes programmable switches with four key modules: “PARSER,” “CS,” “PIT” and “FIB,” and “Encryption and Decryption.” The CS module comprises an index list storing encrypted IDs of cached content and a CS server for storing the actual data, both stored in registers. Upon an INTEREST packet's arrival, the switch checks the index list for a match. If found, it retrieves the corresponding data from the CS server; otherwise, it checks the PIT entry. If no entry exists, it forwards the INTEREST packet following the FIB route, encrypting it before transmission. Upon receiving a DATA packet, it decrypts it and matches the content prefix with the PIT. If a match is found, it caches the content to the CS server and delivers it to a set of users simultaneously.

Centralized-Based Dynamic Content Caching (CDCC)

The paper [86] employed a CDCC caching scheme in which a heuristic algorithm utilized within the SDN controller plays a pivotal role in the efficient selection of cacher groups tasked with storing both popular and time-sensitive content. This method not only optimizes the allocation of caching resources but also enhances the overall performance of content delivery networks within dynamic network environments.

Centrality Content Caching (CCC)

CCC scheme, presented in [87], optimizes content caching across autonomous systems. The controller strategically determines optimal paths and installs them in forwarding devices. When an INTEREST packet reaches a CR, it checks its CS table. If no match is found, the CR consults its CIT for cached content within the autonomous system. If found, the request is directed based on the FIB route. Otherwise, it's forwarded to the controller, which verifies its CIT table for content information across systems. The controller sends the content if a match is found; otherwise, the request is redirected to content producers.

ICC

ICC is a caching approach, presented in [88], designed for DONA, focusing on inter-domain cache cooperation, contrasting with intra-domain cache cooperation. In DONA, RHs can cache incoming DATA packets but don't share availability information with other RHs. ICC enhances cached content availability by having RHs advertise this information to their peering domains.

LCC

LCC, detailed in [89], is a hash-based caching mechanism for inter-domain cooperation. It hashes content names to unique content routers within a domain, effectively mapping content to selected routers for caching. A redundancy control scheme is introduced wherein domains coordinate with neighbors to determine cached content preferences.

Intra-AS Co

Intra-AS Co aims to minimize caching redundancy by enabling neighboring content routers to access cached content defined in [90]. Each content router maintains a Local Cache Content Table (LCCT) and an Exchange Cache Content Table (ECCT). The LCCT records content chunks cached locally, while the ECCT records content cached by neighboring routers. Additionally, content routers periodically share their local cache content information with directly connected neighbors.

Distributed Cache Management (DCM)

The purpose of DCM in work [91] is to maximize the traffic volume served from the caches. By DCM principles, a CM is installed to facilitate the exchange of cache states with other CMs. The determination of cached content placement and replacement depends upon the distributed online cache management algorithm employed by the CM.

CATT

CATT caches a Data packet at a single node along the downloading path within every network as discussed in [92]. Additionally, CATT introduces the content's expected quality to aid content routers in forwarding Interest packets. When content is cached or stored at a content router, its expected quality at that router is considered. Conversely, if the content is not cached at a content router, the expected quality at that router is Hop+1 calculated.

Caching Frameworks Based on ICN Networking

Polyzos et al. [97] present Publisher-Subscriber-Integration for IoT (PSI4IoT), an architecture based on the ICN paradigm, designed to enhance data management and security. The architecture employs an Access Control Provider (ACP) framework to ensure secure authentication and authorization, allowing only authorized users to access protected data. PSI4IoT integrates Proxy Re-Encryption (PRE) for secure data storage, where the device encrypts the data, which is stored with a designated publisher, and a Key Server re-encrypts it for the subscriber, ensuring the device's private key remains secure. Additionally, Hierarchical Identity-Based Encryption (HIBE) guarantees the integrity and verifiability of cached data. Together, these mechanisms optimize content caching, reduce latency, and improve the overall efficiency of IoT ecosystems.

Kurita et al. [98] proposed the Keyword-Based Content Retrieval (KBCR) architecture, which enhances the ICN paradigm for IoT applications. KBCR uses caching nodes to distribute requests based on keywords, initiating a search for relevant data across the network. Multiple responses are then aggregated into a single response, reducing redundant message transmission and improving content retrieval efficiency. This method minimizes latency, reduces network congestion, and enhances performance, making it ideal for resource-constrained IoT systems. Unlike NDN, which requires multiple request messages, KBCR sends a single request that is distributed and merged at intermediate nodes, significantly reducing the number of messages exchanged. Simulations show that KBCR outperforms NDN in large-scale networks, improving response time and scalability while requiring minimal additional processing for message merging.

Song et al. [99] proposed the Smart Collaborative Caching (SCC) scheme to improve the efficiency of information-centric IoT systems by tackling key challenges such as content storage, resource sharing, and node localization. This approach leverages the capabilities of ICN to streamline processes like cluster composition, content management, and position tracking through core actions, including caching, finding, joining, and leaving, all supported by specialized algorithms. Comparative evaluations against traditional methods, such as IP networks and ICN flooding, highlighted SCC's ability to significantly lower packet volume and transmission latency, especially in larger networks. Unlike ICN flooding, which generates excessive traffic and inefficiencies, SCC offers a more controlled and effective solution, reducing congestion and enhancing content retrieval. These advantages position SCC as a highly effective approach for scalable IoT applications.

Hua et al. [100] propose a novel caching design scheme characterized by three key innovative attributes. First, it introduces a Cluster-Based Approach (CBA), which leverages both in-network and end-user devices to cache content closer to the network edge, thereby optimizing content delivery as demand increases. Second, the scheme employs a Near-Path Approach (NPA), strategically positioning caching nodes along delivery paths to minimize latency and enhance retrieval efficiency. Finally, it integrates Reactive and Proactive Caching techniques (RPC) that work together to alleviate network congestion during peak periods. Simulations demonstrate that this design significantly enhances internal link load and path stretch metrics, outperforming eight popular benchmark techniques.

Khodaparas et al. [101] introduce a Pull-Based Caching Scheme (PBCS) aimed at improving content retrieval efficiency in IoT environments. In this method, nodes process data, extract key characteristics, and store the aggregated results locally via the cluster head (CH), instead of directly relaying data to cloud servers. The cached content is kept until its lifetime expires or a request is made, thereby reducing unnecessary transmissions and minimizing power consumption. By enabling local retrieval, the scheme significantly lowers latency compared to fetching content from the cloud. Simulation results show that PBCS increases cache hit rate by 40%, reduces retrieval time, lowers energy consumption, and enhances response times.

Ali Naeem et al. [102] introduce the Most Interested Content Caching (MICC) strategy for NDN-based IoT applications, aimed at improving content dissemination. MICC employs a time-based caching approach and utilizes the intuitionistic fuzzy mode (IFM) to dynamically adjust caching locations based on content interest frequency. Each router tracks content interest, optimizing caching performance metrics such as cache hit ratio and content eviction ratio. Simulations demonstrate that MICC outperforms other strategies like PCS, CCAC, and MPC in content retrieval efficiency, making it a promising solution for enhancing IoT and smart city performance.

Abkenar et al. [103] propose the Distributed Smart Clustering Caching Protocol (DSCCP) to enhance QoS in cache-enabled IoT networks. The protocol employs a smart clustering mechanism that enables Fog Nodes (FNs) to cache content locally based on channel conditions, reducing reliance on Central Nodes (CNs). By selecting optimal destination nodes for processing requests within service delay constraints, DSCCP effectively balances network delay, energy consumption, and overall benefit. Its novelty lies in maximizing network efficiency through intelligent clustering and local caching, outperforming traditional caching methods in responsiveness and resource optimization.

Banerjee et al. [104] proposed the ENROUTE algorithm, which integrates entropy-based techniques with caching to enhance content distribution in ICN. The algorithm uses entropy (EA) as a measure of link congestion and the condition of on-path caches, thereby optimizing routing and caching decisions. Nodes with higher entropy values store more popular content, while links with higher entropy are considered more congested. By improving cache entropy and selecting paths with lower entropy, the algorithm reduces delays and increases throughput. Simulation results demonstrate improvements in delay, cache hit rate, and throughput. Future research will focus on further refining content routing and caching strategies using entropy.

Caching Approaches Based on Machine Learning

The integration of ML-based models in IoT environments leverages predictive analytics to intelligently manage data storage and retrieval, optimizing resource utilization, enhancing system responsiveness, and enabling efficient real-time decision-making. Xiaoming et al. [105] present a Reinforcement Learning (RL) model specifically designed for edge-enabled IoT environments. This model addresses key challenges such as transmission latency and storage costs, aiming to maximize the Quality of Experience (QoE) by balancing content-centric caching quality with user satisfaction. Through its intelligent design, the model significantly improves resource utilization, system responsiveness, and real-time decision-making, demonstrating its potential to enhance IoT operations.

Wu et al. [106] propose a dynamic caching strategy that utilizes user request queues to optimize content delivery. They introduce a queue-aware cache update scheduling algorithm based on the Markov Decision Process (MDP), which aims to minimize the average Age of Information (AoI) delivered to users while accounting for content update costs. Additionally, they present a low-complexity suboptimal scheduling algorithm. Simulation results demonstrate that their approach significantly reduces the average AoI compared to existing strategies that do not incorporate user request queues.

Wang et al. [107] introduced the Federated Learning-based Cooperative Edge Caching (FADE) framework for IoT applications as a solution. This framework aims to enable fast training by decoupling the learning process from cloud-stored data in a distributed-centralized manner, allowing data training to occur on local User Equipment. By incorporating intelligent caching strategies, FADE optimizes the storage and retrieval of frequently accessed data at the network edge, reducing latency and improving response times. In terms of security and privacy, FADE ensures that sensitive data remains on local devices, minimizing data exposure to centralized servers, and leveraging federated learning to preserve user confidentiality. Trace-driven simulations show that FADE significantly outperforms the centralized DRL algorithm, demonstrating notable performance improvements.

Liu et al. [108] introduce an Auto-Regressive (AR) model tailored for IoT environments, supported by detailed mathematical derivations to highlight its effectiveness. By employing the least squares method, they develop a content popularity prediction model that accurately tracks trends in content popularity, enabling more efficient caching strategies in IoT networks. Simulation results confirm the model's effectiveness, showing notable improvements in cache hit ratios, reduced network traffic, and decreased content retrieval delays.

Caching Approaches Based on Deep Learning

The development of DL-based caching solutions for IoT represents a major advancement in optimizing caching efficiency and resource utilization within IoT environments. Zhang et al. [109] present an innovative approach that combines Deep Reinforcement Learning (DRL) with cooperative edge caching to address challenges related to varying data item sizes. The proposed method utilizes the Multi-Agent Actor-Critic (MAAC) algorithm to facilitate collaboration among edge servers, optimizing the caching process. This collaboration enables edge servers to efficiently manage and distribute content, resulting in improved cache hit rates and overall system performance. Simulation results demonstrate the approach's effectiveness, with significant improvements in cache hit rates and system efficiency compared to traditional caching algorithms.

Xinyue et al. [110] introduce Raven, an innovative learning-based caching framework inspired by the optimal Belady algorithm. Raven employs a Mixture Density Network (MDN) to learn the distribution of object arrival times without relying on prior assumptions, allowing for a probabilistic approximation of Belady's algorithm. The framework adapts its cache eviction decisions to the unpredictable and dynamic nature of object arrivals, taking into account factors such as randomness, time-dependent fluctuations, and the absence of fixed arrival patterns. Simulation results demonstrate that Raven outperforms existing algorithms by improving object hit ratios, reducing average access latency, and minimizing traffic to origin servers.

Tian et al. [111] introduce DIMA, a Distributed Intelligent Microservice Allocation scheme. DIMA equips each IoT device with a DRL agent, specifically a D3QN agent, along with a replay buffer. This setup enables the devices to make independent decisions on optimal caching and microservice replacement through interactions with the environment. Simulations demonstrate the effectiveness of DIMA, highlighting its superiority over existing baseline approaches. This innovative solution shows great potential for optimizing microservice caching in dynamic environments, leading to improved performance and resource utilization in IoT ecosystems.

Xu et al. [112] propose a Deep Q-Learning (DQL)-based approach in a software-defined model that integrates computing and caching functionalities. In this model, controllers manage resources from ICN routers and Mobile Edge Computing (MEC) gateways. Traditional algorithms struggle with the high complexity and dimensionality of the problem, as well as the challenge of balancing immediate and long-term rewards. The DQL-based joint optimization approach outperforms other schemes by increasing rewards and efficiently managing resource scheduling, even in dynamic network conditions. It demonstrates superior convergence, faster loss reduction, and more effective decision-making compared to methods that focus solely on caching or computing resources, highlighting the advantages of DQL within the IC-IoT framework.

Ma et al. [113] propose a dynamic caching scheme that leverages Neural Networks (NNs) to optimize content updates in Heterogeneous Networks. The scheme incorporates a queue-aware cache content update scheduling algorithm designed to minimize the average AoI for dynamic content delivery. By utilizing user request queue data, the method predicts and prioritizes the delivery of various dynamic content, ensuring timely and efficient updates. Unlike existing algorithms, the proposed framework anticipates user demands with greater accuracy, thereby improving responsiveness and operational efficiency. Comprehensive evaluations demonstrate that this approach significantly reduces AoI and enhances the overall performance of dynamic content caching systems.

Zhang et al. [114] introduce a dynamic content importance-based caching scheme (D-CICS) that selects and replaces content based on factors like modal type, popularity, size, and network conditions. Unlike traditional methods that focus solely on popularity, D-CICS uses a real-time evaluation model based on Deep Reinforcement Learning with Double Q-Learning (D3QN) to assess the importance of multi-modal content. The scheme incorporates lightweight caching and replacement policies that are specifically tailored for dynamic networks. By efficiently replacing and evaluating content, this approach reduces the risk of unauthorized access or misuse by minimizing unnecessary data transfers and ensuring that only relevant content is cached. This helps mitigate the exposure of sensitive information, thereby enhancing both the security and privacy of the system by limiting the exposure to authorized content only.

Caching Approaches Based on Optimization Applications

Weerasinghe et al. [115] introduced ACOCA, an optimization approach specifically designed for IoT ecosystems to maximize the cost and performance efficiency of Context Management Platforms (CMPs). ACOCA addresses the challenge of efficiently selecting context for caching while managing the additional costs associated with context management. It employs a scalable and selective agent for caching context, implemented using the Twin Delayed Deep Deterministic Policy Gradient (TD3) method. Additionally, ACOCA integrates adaptive context-refresh switching and time-aware eviction policies. A comparative analysis demonstrates ACOCA's superior performance in both cost and performance efficiency, establishing it as a promising optimization solution for IoT ecosystems (Figure 8).

Caching Frameworks Based on ICN Networking

Zhang et al. [116] present a content caching strategy for the IoV that uses Heterogeneous Information Networks (HIN) to address data complexity and the rising demand for popular content. Their method predicts user interest, enabling dynamic caching that reduces network traffic and improves both QoS and QoE. Experiments demonstrate that their approach outperforms traditional methods in content prediction, user satisfaction, and network load reduction, offering better performance, scalability, and lower delays while minimizing reliance on edge computing.

Yang et al. [119] propose an NDN-based framework for CV applications, addressing scalability issues caused by rapid mobility and a large number of vehicles. They introduce the hierarchical hyperbolic NDN backbone (H2NDN), which organizes geographic locations into a hierarchical namespace and router topology. Within this hierarchy, traffic is redirected to higher-level routers, optimizing Interest packet forwarding with static FIB entries. These features significantly improve the efficiency and scalability of the H2NDN backbone, making it well-suited for CV applications.

Chen et al. [120] address the limitations of traditional TCP/IP stacks in IoV scenarios by focusing on spatial-temporal characteristics for data caching in Vehicular Named Data Networking (VNDN). They classify data into emergency safety, traffic efficiency, and entertainment messages, using these categories to determine caching decisions. Their expanded simulation platform shows that their caching strategy outperforms others in terms of hit ratio, hop count, and cache replacement times.

Chen et al. [121] enhance the security of the IoV system by leveraging Named Data Networking (NDN) and Blockchain technology. They utilize Communicating Sequential Processes (CSP) to model the system, ensuring critical properties such as deadlock avoidance, data availability, and protection against fake PIT deletions and Content Store (CS) pollution. These properties are rigorously validated through the Process Analysis Toolkit (PAT). To further reinforce security, the authors integrate Blockchain technology to safeguard the system against unauthorized modifications, ensuring data integrity through immutable transactions. This Blockchain integration fortifies the system against security breaches by providing a tamper-resistant layer that guarantees the authenticity of data exchanges within the NDN-based IoV architecture, enhancing the overall system's resilience to malicious attacks.

Saleem et al. [122] address challenges in VANETs, including communication reliability, security, and scalability. They propose a blockchain-based solution for secure V2V communication, referred to as BC-V2V. This method utilizes a K-Means dynamic clustering algorithm for cluster formation and an NDN approach for efficient content placement, ensuring secure and reliable content distribution. The blockchain framework protects the data exchanged between vehicles, preventing unauthorized access and ensuring the integrity of communications, making it a secure and reliable solution for content distribution in VANETs.

Amadeo et al. [123] propose the Freshness and Popularity Caching (FP-Caching) method to address the challenges of caching transient content in VNDN. This approach focuses on real-time data, such as road traffic updates and parking availability, which quickly lose relevance. While freshness is crucial for caching decisions, especially in safety and non-safety applications, the authors emphasize the importance of considering content popularity, particularly in cases of skewed distribution, such as Zipf-like behavior. This dual focus on freshness and popularity is key to optimizing caching strategies and improving VANET performance.

Zafar et al. [124] propose a hierarchical context-aware content-naming (CACN) scheme for NDN-based VANETs, enabling efficient content forwarding and caching using information from content names. The scheme simplifies communication and storage through an innovative coding technique and reduces latency with a decentralized context-aware notification (DCN) protocol. The DCN protocol broadcasts event notifications, prioritizing vehicles based on location and proximity. Simulations show the scheme's effectiveness in improving VANET communication.

Javed et al. [125] propose a solution for managing large-scale data generation and dissemination in IoV by introducing Vehicular Edge Computing (VEC). In this framework, Road Side Units (RLUs) act as edge servers for caching and task offloading, addressing the increased demands of vehicular applications. The approach includes a task-based architecture for content caching, focusing on content popularity prediction, cache placement, and retrieval.

Ismail et al. [126] propose a proactive caching scheme to reduce latency and improve service quality by utilizing local caches at both the vehicle and Roadside Unit (RLU) layers. This scheme uses high-rate data transmission in vehicular networks to enhance caching efficiency and minimize data costs. The study explores three scenarios: non-cooperative vehicles, cooperative vehicles, and RLU-covered areas, formulating caching placement as an optimization problem to minimize retrieval time.

Caching Approaches Based on Deep Learning

DL has emerged as a powerful solution to address these challenges. IoV applications heavily rely on caching, AI in edge networking, and computing to enhance their functionalities. To tackle these challenges, Li et al. [127] present a study introducing an ICN framework tailored specifically for IoV applications. This framework incorporates a novel deep Q-Learning (DQL) algorithm designed to meet user requirements and optimize caching outcomes, thus enhancing edge service efficiencies. The proposed algorithm shows promising results, contributing significantly to improved environmental learning within the IoV context.

Wang et al. [128] introduce the Cooperative Caching Strategy with Content Request Prediction (CCCRP) to reduce data traffic and improve content request responses in the IoV. The strategy clusters vehicles using the K-means method and predicts content requests using Long Short-Term Memory (LSTM) networks based on historical data. A reinforcement learning method is then applied to optimize caching decisions, enhancing the Quality of Service (QoS) for vehicle requests.

Xu et al. [129] emphasize the need for efficient content processing in the IoV context, particularly for smart city applications. They suggest using Edge Servers (ESs) to ensure quick service delivery and minimize latency for vehicular applications. However, repetitive processing of the same content for multiple users leads to resource inefficiency. To address this, they propose an edge content caching framework for smart cities, integrating RSUs and BSs with ESs. This framework uses a deep learning model, the spatiotemporal residual network (ST-ResNet), to predict traffic flow and apply probability theory to forecast future service demands.

Xu et al. [130] address the cooperative content caching problem using a Constrained Markov Decision Process (CMDP), focusing on content transmission mode and cache replacement to minimize latency. The approach employs Multi-Agent Deep Deterministic Policy Gradient (MADDPG), a Graph Attention Network (GAT)-based aggregator, and a hard-attention collaborator selector to optimize caching. Experimental results show that GAMA-cache outperforms other algorithms, improving service latency and content delivery success rates in real-world datasets.

Caching Approaches Based on Machine Learning

Yu et al. [156] propose the “P2P-FL” federated learning framework for predicting content popularity in the IoV, where vehicles collaboratively update machine learning models. They introduce a dual-weighted model aggregation scheme to minimize the impact of slow vehicles, improving model accuracy. To address privacy, they incorporate a variational autoencoder, which anonymizes data and ensures that only aggregated model updates are shared, preserving confidentiality while enhancing content prediction performance.

Kaci et al. [131] present ROOF-based Named Data Vehicular Networking (NDN) tailored for IoV, featuring intelligent Named Data Caching (iNDC). This machine-learning-based technique predicts content request rates and optimizes caching by retaining popular content on roadside units. iNDC also forecasts storage capacity needs for these units. Performance evaluations reveal that linear and ridge regressions excel at content popularity prediction, while k-Nearest Neighbors offer superior accuracy in predicting the capacity of new roadside units.

Herouala et al. [132] introduce the CaDaCa framework, which leverages data classification techniques. By analyzing real-time user browsing behavior, CaDaCa enhances caching decisions, leading to improvements in cache hit ratio, hop reduction, and overall caching efficiency. In comparison to traditional caching algorithms, CaDaCa exhibits superior performance. Its dynamic adaptation to changing user traffic patterns through data classification ensures continuous optimization of cache management, resulting in more efficient content retrieval and reduced network overhead in NDN-based environments.

Dutta et al. [140] enhance ICN performance in vehicular networks by tackling content availability and redundancy. They use the Monte Carlo Tree (MCT) algorithm for data retrieval and evaluate caching schemes like Level Copy Down and Cache Less for More. Performance evaluations in the ns-3 simulation environment demonstrate improvements in latency, hit ratio, throughput, and overhead metrics, showing the effectiveness of their approach in optimizing ICN caching strategies for better network efficiency in vehicular environments.

Caching Techniques Based on Optimization Applications

Gupta et al. [133] propose an innovative caching strategy to address challenges in vehicular networks, such as limited cache capacity and secure data transfer. Their approach uses a classified scheme, where selected edge nodes store data, and employs a two-level design with the Weighted Clustering Algorithm (WCA) to choose cache heads. Non-Negative Matrix Factorization (NMF) predicts consumer preferences for real user approval on edge nodes. Implemented in Matlab, the method improves cache hit ratio, reduces network delay, and minimizes data loss, making it suitable for the dynamic nature of vehicular networks.

Lam et al. [134] integrate software-defined edge techniques with Unmanned Aerial Vehicles (UAVs) for efficient multimedia content delivery, focusing on caching, resource allocation, multicasting, and 3D trajectory optimization. A service time-aware caching and power allocation (TCP) problem is formally defined and solved using modified genetic algorithms (mGA) to optimize content caching strategies, power distribution for multicast transmissions, and UAV waypoint planning, subject to strict constraints on storage, power, and service time. Simulation results demonstrate that the TCP method significantly enhances network throughput, minimizes service latency, and improves resource utilization efficiency in UAV-assisted content delivery systems (Figure 10).

Caching Approaches Based on Machine Learning

Chen et al. [138] propose an adaptable model for predicting content popularity and modeling content requests by utilizing network features to capture content similarity through a multilevel probabilistic model. The method incorporates Bayesian Learning (BL) to infer model properties even with limited request data. This approach significantly improves content delivery, achieving an impressive 90% hit ratio in a case study focused on video prefetching. The work highlights the synergy between AI and MEC, contributing to optimized content delivery and enhanced user experience.

Sethumurugan et al. [139] propose a machine learning-based cache replacement strategy for MEC, utilizing RL to train a last-level cache (LLC) replacement policy. The approach leverages readily available LLC features without modifying the processor's control and data paths. The trained policy is analyzed with domain knowledge to optimize key features and interactions. The resulting Reinforcement Learned Replacement policy outperforms traditional methods like LRU and DRRIP in benchmark tests.

Dutta et al. [140] propose a method to improve content retrieval in ICN by predicting cache locations for Interest packets. They utilize a Support Vector Machine (SVM) algorithm to predict cache locations based on a simulation dataset. The prediction accuracy of different kernel methods is compared, and simulation results validate the effectiveness of the proposed approach. This study underscores the potential of machine learning in optimizing content retrieval strategies in ICN environments, demonstrating promising results for enhancing retrieval efficiency.

Caching Approaches Based on Deep Learning

Guo et al. [141] propose the QoS-DQN method, which integrates Quality of Service (QoS) optimization with deep-Q-network (DQN) for dynamic resource allocation and edge caching. This approach efficiently groups multimedia application packets into QoS flows, enabling optimized content caching at the network edge for low-latency delivery. What makes this method particularly adaptable is the use of reinforcement learning via DQN, which allows the system to adjust in real time to changes in network conditions and user demands. As mobile devices or vehicles shift between different network cells, causing fluctuations in network topology, QoS-DQN can dynamically reallocate resources, ensuring high-demand content is available at the edge. The key advantage of this approach is its ability to learn from past experiences, enabling it to continuously refine its caching decisions based on real-time data, including traffic patterns, content popularity, and network loads.

Wang et al. [142] present the DECOR scheme, which utilizes advanced Deep Reinforcement Learning (DRL) techniques to enhance caching efficiency and boost user experience. Initially, DECOR learns caching policies from a fixed dataset, providing it with an early understanding of content preferences and network behaviors. However, the key feature of DECOR lies in its capacity to dynamically adapt to real-time conditions by framing the cache replacement problem with a focus on minimizing delay. This capability allows the system to continuously adjust and optimize its caching strategies in response to fluctuating user demands and evolving network conditions. The system's ability to modify its caching decisions based on live inputs makes it highly versatile, especially in settings with rapidly changing network topologies, such as IoV or mobile networks, where users and devices are consistently on the move. As the network topology shifts and user demands vary, DECOR's online DRL model enables it to quickly adapt, ensuring that content is always stored at the most effective locations, reducing latency, and boosting overall performance.

Saputra et al. [143] present two proactive cooperative caching techniques under the PCC-Caching framework. The first method relies on a centralized content server to forecast content demands across various Edge Nodes (ENs), allowing for the early caching of content in locations predicted to experience high demand. However, this centralized approach poses challenges in terms of privacy and communication overhead, as it necessitates the aggregation and processing of large datasets from multiple nodes within the network. To overcome these limitations, the second strategy utilizes a Distributed Deep Learning (DDL) framework, where the content server gathers trained models from mobile edge nodes, combining them into a global model that is continually updated and redistributed. This distributed method reduces the amount of communication required and alleviates privacy concerns by keeping data at the local level. Moreover, it improves the system's adaptability to real-time changes in user preferences and network topology by enabling the system to learn and adjust dynamically based on distributed data.

Khanna et al. [144] tackle the issue of optimizing content delivery in Mobile Edge Computing (MEC) systems, which face challenges due to the dynamic nature of user mobility and fluctuating content demand. To address these issues, they propose the Federated Graph Reinforcement Learning (CFGRL) approach, which integrates graph neural networks, federated learning, and DRL. The use of graph neural networks allows the system to better understand the complex interconnections between network nodes, improving the accuracy of caching decisions. Federated learning preserves user privacy by enabling model training on distributed devices without sharing sensitive data, while DRL helps the system adjust its strategies in response to real-time shifts in network conditions and user behavior. By continuously tracking content popularity and adjusting caching strategies accordingly, CFGRL optimizes content delivery, reduces latency, and ensures privacy, making it highly effective in handling rapidly changing network topologies and fluctuating user demands.

Caching Techniques Based on Optimization Applications

Cao et al. [145] introduce the Cooperative MEC Caching (CMCC) approach to enhance caching efficiency by optimizing both computational and communication resources. Their method integrates nearby helper nodes to offload computational tasks, thereby improving caching performance in vehicular networks. By applying joint optimization techniques, they developed an efficient algorithm that delivers optimal solutions, particularly in partial offloading scenarios. Numerical results highlight the benefits of this cooperative scheme, demonstrating significant improvements in MEC performance within vehicular environments.

Ji et al. [146] propose the Hybrid Offloading and Cooperative Energy Efficiency (HO-CEE) method, which includes two offloading schemes using Hybrid Access Points (HAP). The first offloads tasks to the edge cloud for energy optimization via caching, while the second employs cooperative communication, with a nearer user relaying tasks from a farther user, enhancing Energy Efficiency (EE). Simulation results confirm that the cooperative scheme outperforms the noncooperative scheme, significantly improving EE and mitigating the near-far effect.

Kim et al. [147] propose the Mobile Edge Computing Caching System (MECCS), designed to optimize cache hits for popular video content in edge servers. Their approach addresses limitations in Video-on-Demand caching by incorporating a request model that accounts for Short-term Time-Varying (STV) popularity in live streaming. Using an exponential distribution for request rates, it improves cache hit rates and reduces backhaul traffic. Experimental results show MECCS's superiority over existing methods, particularly in live-streaming scenarios.

Li et al. [148] introduce the Delay Control Strategy Joint Service Caching and Task Offloading (DCS-OCTO) algorithm, which facilitates random online service caching without requiring prior knowledge of service popularity. The algorithm optimizes system delay while adhering to energy consumption constraints in MEC systems. Simulation results demonstrate that DCS-OCTO outperforms other algorithms in minimizing delay and energy consumption while considering access point storage and energy limitations.

Zhang et al. [149] propose the OCP-SBS method for large-scale user-centric mobile networks. Their approach jointly optimizes content placement, Small Base Station (SBS) clustering, and bandwidth allocation while considering radio resource constraints and network variability. This method minimizes file transmission delay and enhances network efficiency, achieving delay-optimal SBS clustering for cooperative caching.

Wang et al. [150] introduce the Joint Transmission and Cache (JTC) scheme to address backhaul bandwidth consumption in MEC servers. They consider video popularity and quality in their analysis, integrating Adaptive Bitrate (ABR) technology to dynamically adjust video bitrates based on network conditions. This approach maximizes system revenue while accommodating differentiated video quality requests, effectively balancing video popularity and quality to optimize caching strategies.

Ansari et al. [151] propose a Mobile Edge Internet of Things (MEIoT) framework aimed at enhancing the mobile user experience by optimizing caching. This framework integrates fiber-wireless technology, cloudlets, and software-defined networking to bring computing and storage resources closer to mobile IoT devices. By connecting IoT devices to proxy Virtual Machines (VMs) in nearby cloudlets, it enables real-time data caching and retrieval. To further reduce delays and energy consumption, the framework incorporates dynamic proxy VM migration methods, ensuring more efficient resource utilization and significantly improving overall performance (Figure 12).

Caching Frameworks Based on ICN Networking

Ullah et al. [154] propose Proactive Name-Based Content Prefetching (PNCP), a strategy that integrates ICN with EC to enhance content distribution. By proactively fetching dynamic content and utilizing ICN data structures like the CS, PIT, and FIB within the. NET framework, PNCP improves data locality and reduces network bandwidth demands. Experimental results show PNCP achieves higher cache hit ratios and lower delays, outperforming conventional edge systems and prior ICN approaches.

Wang et al. [155] propose the NDN-based IoT Edge (NIoTE) framework to enhance data communication efficiency in IoT edge environments. NIoTE integrates IoT with edge computing and clustering techniques to address the resource constraints of IoT devices. Through clustering, the framework enables resource sharing, collaborative data collection, request aggregation, and the establishment of the FIB. Its modular architecture ensures seamless integration with existing IoT infrastructures, supporting rapid deployment and scalability across diverse IoT environments. Experimental results confirm the effectiveness of NIoTE, highlighting substantial reductions in data communication latency and costs.

Caching Approaches Based on Machine Learning

Yu et al. [156] propose a Federated Learning-based Proactive Content Caching (FPCC) scheme. This scheme addresses the challenges of predicting file popularity amidst fluctuating trends and limited cache capacity by employing privacy-preserving federated learning, which aggregates user-side updates to build a global model. Using hybrid filtering with a stacked autoencoder (SAE), FPCC identifies and selects popular files for caching. Experimental results demonstrate that FPCC not only enhances cache efficiency compared to traditional algorithms but also ensures robust data privacy.

Fan et al. [157] introduce Popularity-aware cache (PA-Cache), a system designed to adapt to sudden surges in requests, which is essential due to the smaller cache sizes in edge networks. Unlike previous approaches, PA-Cache leverages a comprehensive set of content features to dynamically predict time-varying content popularity. To reduce the computational overhead of conventional DNNs, PA-Cache begins with a shallow network for rapid convergence, later transitioning to more robust DNNs as request volumes increase. This approach ensures both scalability and computational efficiency.

Cui et al. [158] propose CREAT, a system that integrates IoT devices, edge nodes, remote cloud, and blockchain to improve data processing, caching, and security in IoT environments. CREAT uses federated learning to allow edge nodes to train models locally without uploading sensitive data. To address communication bottlenecks, it employs techniques like K-Means for identifying important gradients and clustering-based quantization to reduce transmission data volume. Experimental results show that CREAT enhances cache hit rates and reduces data upload time, improving overall content caching efficiency.

Caching Approaches Based on Deep Learning

Ndikumana et al. [159] propose Deep Infotainment Caching (DIC), a deep learning-based approach for efficient edge caching in self-driving cars. DIC predicts cached content and vehicle proximity to multi-access edge computing (MEC) servers. The method includes content prediction models, a communication model for retrieval, a caching model, a computation model for serving content, and an optimization problem to minimize downloading delays. Simulation results show a 97.82% prediction accuracy for cached content and highlight significant delay reduction.

Kang et al. [160] propose Priority-Based Optimal Cache Placement (POCP), an ICN caching scheme that integrates content popularity and priority. Using a seq2seq LSTM model, POCP predicts content requests and communicates these predictions to the SDN controller, which optimizes content placement in MECs and core routers. This minimizes delivery costs while ensuring high QoE for prioritized content. Performance evaluations show POCP outperforms traditional caching schemes in terms of QoE satisfaction, cost efficiency, and overall effectiveness.

Zhu et al. [161] introduces the Edge-Graph Cache Replacement (EGCR) method, which utilizes a graph-based representation of cache access data and incorporates Graph Neural Networks (GNNs) for cache replacement, enabling dynamic adaptation to complex access patterns in real-time. GNNs model the relationships between cache data and access sequences as graph structures, capturing both spatial and temporal dependencies within the data. This enables the system to optimize cache decisions by learning from intricate patterns in access behavior. The approach significantly improves the Cache Hit Ratio, adapts to fluctuating workloads, and outperforms conventional cache replacement strategies.

Caching Techniques Based on Optimization Applications

Liu et al. [162] propose the Optimized Data Caching Model (ODCM), which integrates the benefits and costs of data caching while considering varying data popularity. The model formulates the caching problem in edge computing as an integer programming (IP) problem and uses the Page-Hinckley-Test (PHT) to assess data popularity. An approximation approach is introduced to efficiently find near-optimal solutions, aiming to optimize the service provider's revenue, especially in large-scale scenarios, with a detailed analysis of the approximation ratio.

Xia et al. [163] propose a solution to the collaborative edge data caching (CEDC) problem in edge computing environments, where edge servers at base stations face resource constraints. They model the CEDC problem as a constrained optimization issue from the app vendor's perspective and prove its NP-completeness. To address this, they introduce an online algorithm, CEDC-O, with proven performance bounds. The algorithm efficiently solves the CEDC problem across multiple time slots without requiring future information. Extensive simulations using real-world data show the algorithm's effectiveness in optimizing data caching in edge computing environments.

Xia et al. [164] address the NP-completeness of the Edge Data Caching (EDC) problem with two approaches: IPEDC, an optimal solution using Integer Programming for precise caching strategies, and AEDC, an approximation method for large-scale scenarios that balances solution quality and efficiency. Their extensive experiments show that IPEDC achieves optimal solutions, while AEDC efficiently handles large-scale challenges, outperforming four representative methods.

Xia et al. [165] introduce the Data, User, and Power Allocation (DUPA3) problem, aiming to optimize data, user, and transmit power allocations to maximize data rates and serve more users. They prove the NP-completeness of the problem and present a potential game framework, showing the existence of a Nash equilibrium. To solve the DUPA3 problem, they propose a decentralized algorithm, DUPA3Game, and evaluate its performance through theoretical and experimental analysis. This work lays the foundation for efficient resource allocation in MEC environments, ensuring fast data retrieval for users.

Gao et al. [166] propose a novel approach for efficient content caching and service placement in edge computing systems. They introduce a class of stochastic models, using Layered Queueing Networks (LQNs) to optimize service placement, followed by an enhanced Joint Caching and Service Placement (JCSP) model that incorporates caching components. Extensive simulations with real-world data show that JCSP outperforms baseline methods, offering a superior balance between system response time and memory consumption (Figure 14).

Caching Frameworks Based on ICN-Networking

In the realm of Fog computing, two notable studies emphasize the effectiveness of ICN caching strategies. Ammar et al. [169] introduce 3Q, a low-complexity caching strategy seamlessly integrated into an ICN-based Edge/Fog architecture. This strategy focuses on efficiently managing service instances and source codes while minimizing overhead, achieving performance levels comparable to optimal strategies. Similarly, Alghamdi et al. [170] propose an innovative Content-Aware CDN (CA-CDN) architecture within Fog computing that leverages ICN principles to optimize content delivery. Their approach integrates a content-aware routing protocol, resulting in significant performance improvements over alternative methods. Figure 16 presents the classifications of caching in Fog Computing.

Ibrahim et al. [171] propose the Fog-based Browser (FBB) framework, which enhances QoS and reduces latency by positioning cloud capabilities closer to end devices. Their approach demonstrates significantly faster web page load times, with experiments showing latency reductions of up to 85%, thus improving the overall user experience. To address security concerns, the FBB framework likely incorporates secure communication protocols, such as encryption for data transmission between edge devices and the cloud, ensuring the confidentiality and integrity of sensitive data.

Hassan et al. [172] propose the Vehicular Multi-layer Caching (VMC) system, which enhances content sharing among vehicles in urban areas by integrating data centers, fog controllers, and edge nodes. Their transient caching approach dynamically adjusts cache offloading based on traffic intensity, optimizing content availability across various fog configurations. To ensure security and privacy, the system uses encryption for secure content transmission and implements access control to restrict data retrieval and storage to authorized users. By relying on localized caching, it minimizes data transfer, reducing the exposure of sensitive vehicle and user information.

Caching Techniques Based on Optimization Applications

Shahid et al. [173] introduce the Energy-Efficient Fog Caching (EEFC) approach, which optimizes content distribution in Fog networks by categorizing Fog nodes into clusters and employing caching strategies based on energy and capacity. Gateways manage content requests, forwarding them to the appropriate clusters, while active nodes are selected dynamically. Popular content is cached efficiently, and load balancing algorithms ensure equitable workload distribution. Their approach demonstrates significant reductions in energy consumption and latency compared to traditional methods.

Talaat et al. [174] introduce the Effective Cache Replacement Strategy (ECRS), a load balancing technique for real-time Fog Computing environments. ECRS enhances cache utilization by integrating optimization-based prefetching policies and organizing Fog nodes into regions managed by master nodes. Using a reactive routing protocol and a table-driven strategy, ECRS improves data access, maximizing cache hits and reducing access latency, thereby enhancing Fog computing performance.

Caching Approaches Based on Deep Learning

Bhandari et al. [175] developed the Deep Learning Caching Control (DLCC) algorithm to predict content popularity across various categorical classes. Using a 2D CNN model trained on a 1D dataset, the algorithm effectively forecasts the popularity of future data. Based on these predictions, DLCC implements a caching strategy that stores data from the most popular class in the cache memory of nearby Fog Access Points (F-APs). The performance of this caching policy is evaluated using metrics such as DL accuracy, cache-hit ratio, and system delay, demonstrating its effectiveness in optimizing cache resource utilization and reducing latency.

Hasan et al. [176] introduce an innovative content management and placement framework that utilizes Quantum Memory Modules (QMM) for storage. Their Deep Learning-Associated Quantum Computing (DLAQC) framework integrates a DL agent at the network edge with QMM, employing the Self-Organizing Maps (SOMs) algorithm to prioritize and effectively store cached content. Similarly, Huang et al. [177] propose a D2D-assisted fog computing architecture that integrates FL and DNN to adapt to user preferences, aiming to reduce content fetch delays and ensure QoS. They introduce the Cluster-based User Preference Estimation (CUPE) algorithm for optimized content caching placement, leveraging deep learning to enhance caching efficiency and system performance.

Caching Approaches Based on Machine Learning

Wu et al. [178] propose a federated learning-based approach to predict content popularity, enhancing content caching in fog computing by incorporating user preferences and contextual data. Their method utilizes the SVRG algorithm for local model training and combines it with DANE for constructing a global model, improving content prediction while reducing computational overhead. In terms of security and privacy, the method leverages federated learning to keep sensitive user data on local devices, avoiding transmission of raw data over the network. Only aggregated model updates are shared, which protects user privacy by reducing the risk of exposing personal information and ensuring secure and efficient content caching.