PaperlayerView exact revision on arXiv

Source-linked AI summary

A Survey on Semantic Communication Networks: Architecture, Security, and Privacy

Shaolong Guo, Yuntao Wang, Ning Zhang, Zhou Su, Tom H. Luan, Zhiyi Tian, Xuemin, Shen

arXiv:2405.01221v2cs.NI

TL;DR

SemComNet’s multi-agent fundamentals and security and privacy aspects had not been comprehensively surveyed. This paper surveys the field, introduces a three-layer architecture, and reviews threats and countermeasures across its functional layers.

  • Problem

    SemComNet’s multi-agent fundamentals and security and privacy aspects had not been comprehensively surveyed.

    Source: C. Related Works · p. 3

  • Method

    The survey presents a three-layer SemComNet architecture and reviews working modes, technologies, threats, countermeasures, and future directions.

    Source: A. Architecture of SemComNet · p. 5, §VI · p. 28

  • Results

    The survey identifies security and privacy threats across SemComNet’s three functional layers and reviews countermeasures.

    Source: §VI · p. 28

  • Takeaways & Limitations

    The survey highlights paths and methods intended to foster secure SemComNet for intelligent applications.

    Source: C. Related Works · p. 3

Abstract

from arXiv · show

With the rapid advancement and deployment of intelligent agents and artificial general intelligence (AGI), a fundamental challenge for future networks is enabling efficient communications among agents. Unlike traditional human-centric, data-driven communication networks, the primary goal of agent-based communication is to facilitate coordination among agents. Therefore, task comprehension and collaboration become the key objectives of communications, rather than data synchronization. Semantic communication (SemCom) aims to align information and knowledge among agents to expedite task comprehension. While significant research has been conducted on SemCom for two-agent systems, the development of semantic communication networks (SemComNet) for multi-agent systems remains largely unexplored. In this paper, we provide a comprehensive and up-to-date survey of SemComNet, focusing on their fundamentals, security, and privacy aspects. We introduce a novel three-layer architecture for multi-agent interaction, comprising the control layer, semantic transmission layer, and cognitive sensing layer. We explore working modes and enabling technologies, and present a taxonomy of security and privacy threats, along with state-of-the-art defense mechanisms. Finally, we outline future research directions, paving the way toward intelligent, robust, and energy-efficient SemComNet. This survey represents the first comprehensive analysis of SemComNet, offering detailed insights into its core principles as well as associated security and privacy challenges.

I. INTRODUCTION

As AI-enabled agents become increasingly connected, communication must support their task execution and coordination. Semantic communication transmits task-relevant meaning, but extending paired systems to multiple agents raises scaling, knowledge-consensus, and topology challenges.

I. INTRODUCTION

  • By 2030, over 125 billion AI devices, termed agents, are predicted to connect to the Internet.The introduction frames this growth as part of a shift from connected things to connected intelligence.Source: §I · p. 1

A. Motivation of Semantic Communication

  • SemCom identifies agents’ computing requirements and transmits only the semantic meanings needed, using continuously updated knowledge bases to reduce redundant data.This approach is presented as supporting persistent, memory-based communication and collaborative complex tasks.Source: A. Motivation of Semantic Communication · p. 1
  • Paired SemCom addresses two agents, while SemComNet extends semantic-oriented services to multiple agents sharing intents and objectives.The survey also describes paired SemCom as efficient, reliable, and compatible with existing infrastructure.Source: A. Motivation of Semantic Communication · p. 2
  • Scaling SemComNet increases training and deployment demands because semantic models are resource-intensive and typically tailored to particular transceiver pairs and tasks.The passage warns that these demands can overwhelm individual agents as networks grow.Source: A. Motivation of Semantic Communication · p. 2
  • Multi-agent networks must coordinate and update multiple compatible knowledge bases as knowledge changes, while adapting models and consensus to agents joining or leaving.The text says a single agent cannot realistically coordinate these knowledge bases independently.Source: A. Motivation of Semantic Communication · p. 2

B. Motivation of Securing SemComNet

The paper frames SemComNet as a multi-agent extension of paired semantic communication, with security and privacy challenges that prior surveys have not treated comprehensively. It surveys the network’s architecture, operating modes, threats, defenses, and future directions.

B. Motivation of Securing SemComNet

  • SemComNet faces hardware, communication, and private-KB threats, including sponge-example DoS, covert semantic manipulation, and unauthorized retrieval of sensitive information.Semantic attacks can manipulate SI extraction or interpretation through imperceptible data changes; private KBs may contain preferences and behaviors.Source: B. Motivation of Securing SemComNet · pp. 2–3
  • Security responses face interoperability, decentralized oversight, and KB-synchronization challenges in heterogeneous, autonomous, large-scale networks.The paper also notes that desynchronization and KB-poisoning attacks can worsen coordination challenges.Source: B. Motivation of Securing SemComNet · p. 3

C. Related Works

  • This survey addresses gaps in prior work by reviewing SemComNet architecture and modes alongside security, privacy, defenses, and future directions.Earlier surveys largely focused on paired agents or omitted comprehensive treatment of working modes, use cases, and security and privacy.Source: C. Related Works · p. 3
  • The survey presents a taxonomy of security and privacy threats across three layers, reviews defense solutions, and identifies future research directions.It considers solutions from academia and industry and discusses their potential for secure, trustworthy, reliable, privacy-preserving SemComNet.Source: C. Related Works · p. 3

II. AN OVERVIEW OF SEMANTIC COMMUNICATION NETWORKS

  • The overview section proceeds from SemCom’s roadmap and SemComNet motivation to its envisaged architecture, working modes, and enabling technologies.The section introduction describes this progression.Source: §II · p. 4

1) Roadmap of SemCom:

  • SemCom’s roadmap shifts from single-modal to multi-modal transmission and from single-task to multi-task execution, with added complexity and possible task interference.The paper also describes a shift from paired to networked SemCom, which brings training, KB-consensus, and topology-dynamics challenges.Source: 1) Roadmap of SemCom · p. 4
  • The proposed networked paradigm combines KB maintenance, task scheduling, resource allocation, SI delivery, environmental perception, and agent cognition for multi-agent interaction.The paper presents SemComNet as supporting agents that jointly accomplish complex tasks in 6G and beyond applications.Source: 1) Roadmap of SemCom · p. 4
  • Clustered SemCom coordinates agents on a common task, whereas networked SemCom is intended for more complex scenarios spanning multiple transmission tasks.The paper describes networked SemCom as supporting coordination across those tasks.Source: 1) Roadmap of SemCom · p. 4

A. Architecture of SemComNet

The proposed SemComNet architecture organizes multi-agent information sharing into three functional layers. Its control layer manages shared knowledge, task scheduling, and network resources.

A. Architecture of SemComNet

  • The SemComNet architecture has control, semantic transmission, and cognitive sensing layers for multi-agent information sharing.The transmission layer provides semantic-oriented delivery, while cognitive sensing handles environment sensing, intent inference, and contextual knowledge.Source: A. Architecture of SemComNet · p. 5, Figure 3 · p. 6

1) Control Layer:

  • The control layer coordinates knowledge-base maintenance, task scheduling, and resource allocation across the cloud-edge-end architecture.It monitors task requirements and agent capabilities when scheduling work, and allocates communication and computing resources.Source: 1) Control Layer · pp. 5–6, A. Architecture of SemComNet · p. 5
  • Shared knowledge bases provide semantic models and contextual knowledge that can reduce training from scratch and support semantic reasoning and decoding.They may be represented as knowledge graphs, structured databases, or parameters in GAI models.Source: 1) Control Layer · p. 5
  • Cloud resources support global knowledge-base maintenance and model training, while edge resources reduce data-transfer latency and privacy leakage.End agents handle simpler tasks, including sensing, privacy-sensitive knowledge updates, and semantic information delivery.Source: 1) Control Layer · pp. 5–6

2) Semantic Transmission Layer:

The semantic transmission layer prepares agents for communication and then carries out semantic-aware transmission. The cognitive sensing layer gathers environmental information, infers agent intents, and organizes knowledge for the network.

2) Semantic Transmission Layer:

  • Transmission preparation aligns agents with shared knowledge bases and equips them with semantic models through private, federated, or centralized learning.Semantic-aware transmission proceeds through multi-source data fusion, semantic-information extraction, reliable physical-layer transmission, and semantic recovery and enhancement.Source: 2) Semantic Transmission Layer · p. 6
  • Semantic encoders extract desired information from multimodal data, filtering redundancy and adapting its detail to receiver knowledge, task, and channel conditions.High-level information can suit classification, while lower-level detail supports scene comprehension or reconstruction.Source: 2) Semantic Transmission Layer · pp. 6–7
  • Precoding and channel coding support reliable multi-agent transmission; receivers detect signals, decode with shared knowledge, and can use GAI models to correct blurred or missing content.The layer also describes deep-learning joint source-channel coding as an end-to-end approach to efficiency and flexibility.Source: 2) Semantic Transmission Layer · p. 7

3) Cognitive Sensing Layer:

  • Cognitive sensing uses environmental sensors and semantic understanding to turn raw data into descriptions, then infers intents from behavior and interaction histories.The described inference approaches include traditional machine learning, deep learning, reinforcement learning, and cognitive models; they suit different data and adaptation conditions.Source: 3) Cognitive Sensing Layer · p. 7
  • Cognitive sensing structures sensor data and inferred intents into knowledge, including cross-modal knowledge graphs, and coordinates private and shared knowledge-base updates.Agents may hesitate to share private knowledge for privacy reasons; federated learning is presented as a way to build shared knowledge without exposing sensitive information.Source: 3) Cognitive Sensing Layer · pp. 7–8, Figure 5 · p. 8

B. Working Modes of SemComNet

SemComNet working modes range from two-agent interaction to collaboration within and across task-based clusters. The paper also surveys the technologies supporting these modes.

1) Paired Semantic Communication:

  • Paired SemCom exchanges semantic information between two agents after they synchronize background knowledge and align shared knowledge bases.The interaction extracts, transmits, and reconstructs semantic information, then assesses quality and adapts transmission to channel conditions.Source: 1) Paired Semantic Communication · pp. 8–9
  • Knowledge distillation can produce lightweight semantic codecs from larger models for agents with limited computing power.In paired SemCom, codecs are prepared alongside shared-knowledge alignment to reduce agents’ computation burden.Source: 1) Paired Semantic Communication · p. 8

2) Clustered Semantic Communication:

  • Clustered SemCom lets agents collaborate on a common task by sharing semantic fragments, fusing received information, and requesting retransmission of missing content.Semantic fragments are meaningful information units whose definitions and standardization remain evolving.Source: 2) Clustered Semantic Communication · p. 9

3) Networked Semantic Communication:

  • Networked SemCom connects task-based clusters, using semantic translators to bridge knowledge or codec mismatches and support cross-task coordination.Its multi-layered knowledge bases include public, task-specific, and private knowledge.Source: 3) Networked Semantic Communication · p. 9

C. Enabling Technologies of SemComNet

  • Enabling technologies include deep learning, transfer and continual learning, generative models, and reinforcement learning for semantic processing and adaptation.The survey describes generative models as a source of training samples and reinforcement learning as supporting autonomous reasoning and real-time decisions.Source: C. Enabling Technologies of SemComNet · p. 10

2) Networking Technology:

SemComNet combines networking, knowledge representation, and ubiquitous computing to support agent interaction, with use cases spanning autonomous driving, holographic communication, and the Metaverse.

2) Networking Technology:

  • SemComNet uses IoT, BLE, LoRa, and 5G/B5G for reliable, scalable, low-latency SI transmission between agents.BLE supports short-range intra-cluster communication, while LoRa supports long-range inter-cluster links.Source: 2) Networking Technology · p. 10

3) Knowledge Representation Technology:

  • Knowledge graphs, ontologies, and scene graphs organize contextual information and experience into forms agents can manage, search, and reason over.They represent structured knowledge, concepts and relationships, or objects and relationships in images, respectively.Source: 3) Knowledge Representation Technology · p. 10

4) Ubiquitous Computing:

  • Ubiquitous computing embeds resources across agents, while cloud-edge-end infrastructure provides shared-model training, nearby processing, and coordinated KB access.Cloud tiers offer powerful training resources; edge tiers process nearby agents’ tasks rapidly.Source: 4) Ubiquitous Computing · p. 10

5) Summary:

  • AI, networking, knowledge representation, and ubiquitous computing jointly support semantic-aware, intelligent interaction among agents.The paper assigns these technologies roles in semantic understanding and optimization, transmission, knowledge management, and computation.Source: 5) Summary · p. 10

1) Autonomous Driving:

  • SemComNet integrates intent and semantics into autonomous-driving communication to support efficient vehicle-road collaboration under perception-accuracy and latency demands.Vehicles can use shared edge knowledge for local decisions and collaborate through SI exchange; one cited system uses JSCC and a fusion module to reduce redundancy and coordinate tasks.Source: 1) Autonomous Driving · pp. 10–11

2) Holographic-type Communications:

  • Holographic communication’s Tbps and sub-1 ms demands are challenging, while SemComNet can reduce transmitted volume by prioritizing task-relevant semantic information.SemHolo demonstrates real-time interactive telepresence with visual quality and reconstruction fidelity; a shared semantic context can also support multimodal recovery.Source: 2) Holographic-type Communications · p. 11

3) Metaverse:

  • For Metaverse interactions, SemComNet transmits essential semantic content, such as key movements, instead of full 3D gestures, reducing data load and accelerating processing.Large-scale activities also strain server and user-device computing and transmission resources.Source: 3) Metaverse · p. 11

E. Semantic Metrics

Semantic metrics assess information significance, meaning, or both, and can inform security monitoring. The section then surveys control-layer threats involving resource exhaustion, shared knowledge bases, and semantic models.

E. Semantic Metrics

  • Semantic metrics comprise significance-oriented, meaning-oriented, and combined measures, evaluating information value and timeliness, semantic fidelity, or both.Age of information and value of information are significance-oriented examples; semantic QoE combines fidelity, communication costs, and user experience.Source: E. Semantic Metrics · pp. 11–12
  • Monitoring AoI or VoI can help identify and discard excessively old or delayed data that attackers might use to mislead receivers.The cited example is replaying or delaying outdated or irrelevant data to confuse the receiving agent.Source: E. Semantic Metrics · p. 12

III. SECURITY AND PRIVACY THREATS IN SEMCOMNET

  • Control-layer attacks target orchestration, resources, shared KBs, and semantic models, including sponge examples, poisoning, unauthorized access, and model theft.The section frames SemComNet security and privacy around confidentiality, integrity, and availability of data, knowledge, and models.Source: A. Threats to Control Layer in SemComNet · p. 12, §III · p. 12

A. Threats to Control Layer in SemComNet

  • Sponge examples can defeat hardware acceleration in DL components, causing excessive energy use and response delays; worst-case thresholds can reject inputs exceeding natural-example bounds.The cited generation methods are gradient-based white-box and genetic-based black-box approaches.Source: A. Threats to Control Layer in SemComNet · p. 12
  • KB poisoning injects false or harmful knowledge to deceive receivers, while desynchronization can create inconsistent KBs and impair agents’ shared understanding.The section identifies access control and blockchain as possible poisoning defenses, but says desynchronization attacks and defenses remain unexplored.Source: A. Threats to Control Layer in SemComNet · pp. 12–13
  • KB mismatches can cause semantic noise when agents withhold sensitive knowledge or KBs evolve; unauthorized access can expose private information or enable tampering.The section also identifies query misguiding and feature-importance attacks, with defenses against the latter remaining unexplored.Source: A. Threats to Control Layer in SemComNet · pp. 13–14

B. Threats to Semantic Transmission Layer in SemComNet

The survey describes threats to both semantic transmission and cognitive sensing: attacks can manipulate transmitted meanings, exploit feedback or relays, and corrupt sensing, knowledge, or agent behavior.

B. Threats to Semantic Transmission Layer in SemComNet

  • Semantic transmission exposes SI and tasks on broadcast channels while AI-based codecs remain vulnerable to poisoning and adversarial manipulation that can distort meaning.The section covers data and model poisoning, semantic adversarial attacks, jamming, eavesdropping, relay threats, and feedback exploitation.Source: B. Threats to Semantic Transmission… · p. 14
  • Semantic jamming targets meaning rather than bit-level communication, while relays can expose stored private knowledge or alter relayed SI.The described GAN-inspired jammer–receiver framework seeks resilience, but assumes the jammer shares the sender’s network structure; relay defenses remain unexplored.Source: B. Threats to Semantic Transmission… · pp. 14–15
  • Feedback leakage can reveal agent preferences and behavior, while tampering can trigger inappropriate transmission-rate adjustments; defenses remain an open issue.The transmitter may send more detailed, low-level SI when receiver feedback indicates difficulty interpreting content.Source: B. Threats to Semantic Transmission… · p. 15

C. Threats to Cognitive Sensing Layer in SemComNet

  • Cognitive-sensing threats include deceptive information, false sensed data, free-riding, impersonation, trust-management difficulties, and malware affecting data and knowledge.The layer’s sensing, intention inference, knowledge discovery, and private-KB capabilities create vulnerabilities during perception, processing, and sharing.Source: C. Threats to Cognitive Sensing Layer… · p. 15
  • GPS spoofing can insert flawed environmental semantics into KBs, while free-riding withholds useful sensing contributions and can degrade service availability and QoE.Robust authentication and provenance tracking may mitigate false-data injection; trust evaluation and access control are proposed for free-riding.Source: C. Threats to Cognitive Sensing Layer… · p. 15
  • Trust management is difficult when real-time behavior monitoring may compromise confidentiality and agents share incomplete data; malware can also spread and steal or delete information.The survey points to trust-free architectures and efficient trust evaluation, and to combined malware detection, defense, and prevention.Source: C. Threats to Cognitive Sensing Layer… · pp. 16–17

D. Summary and Lessons Learned

SemComNet threats include inherited attacks with heightened, reduced, or altered effects, as well as new KB- and AI-related risks. The survey reviews defenses across five security and privacy aspects, including model-poisoning countermeasures.

D. Summary and Lessons Learned

  • SemComNet threats include amplified relay risks, harder semantic eavesdropping, DoS variants such as sponge examples, and KB-related attacks.Cross-layer dependencies can let cognitive-layer contamination propagate through shared KBs into semantic transmission and impair interpretation.Source: D. Summary and Lessons Learned · p. 17
  • The survey organizes defenses across semantic model security, semantic transmission security, reliability, trust, and data/knowledge security.Its preceding threat taxonomy groups attacks across the control, semantic transmission, and cognitive sensing layers.Source: D. Summary and Lessons Learned · p. 17

IV. SECURITY AND PRIVACY COUNTERMEASURES IN SEMCOMNET

  • The countermeasure discussion covers five areas: semantic model security, semantic transmission security, reliable SemComNet, trust management, and data and knowledge security.The section presents these as defenses tailored to SemComNet.Source: §IV · p. 17

1) Defense Strategies Against Semantic Model Poisoning:

  • For centralized model poisoning, sanitization can detect and mitigate malicious parameter changes; fine-pruning removes suspicious neurons and retrains on clean data.The cited input-perturbation method faces scalability issues for large models and assumes clean training data.Source: 1) Defense Strategies Against Semantic… · p. 17
  • For distributed learning, robust aggregation such as Krum filters abnormal updates, while cross-validation can adjust client update weights using other clients’ local data.Krum’s complexity rises with participant count because it calculates pairwise distances.Source: 1) Defense Strategies Against Semantic… · p. 17

2) Adversarial Training for Semantic Noise Resistant SemComNet:

The survey reviews semantic-model defenses and transmission-security approaches, including adversarial training, sponge-example safeguards, model-IP protection, and semantic-data encryption. It also highlights their limitations, including generalization, resource, and key-management burdens.

2) Adversarial Training for Semantic Noise Resistant SemComNet:

  • Adversarial training improves resilience to semantic adversarial examples, but narrowly crafted examples can raise compute costs and overfitting risks.The cited work uses fast gradient method perturbations and reports high BLEU and BERT scores.Source: 2) Adversarial Training for Semantic… · p. 17
  • SemMixed significantly reduces attack success rates against physical-layer adversarial perturbations in SemComNet.The experiments also evaluate PSNR, SSIM, and misleading rate.Source: 2) Adversarial Training for Semantic… · p. 18

3) Safeguarding SemComNet Against Sponge Examples:

  • Worst-case performance thresholds can reject resource-intensive sponge inputs, but calibration and subthreshold attacks limit threshold-only defense.High volumes of sponge examples may still overwhelm resources and cause denial of service.Source: 3) Safeguarding SemComNet Against… · p. 18

4) IP Protection for Securing Model in SemComNet:

  • DNN watermarks support model-ownership checks, while task-agnostic output watermarks retain extraction success against surrogate models.Blockchain is also presented as a traceable mechanism for sharing semantic codecs and knowledge.Source: 4) IP Protection for Securing Model in… · p. 18

B. Semantic Transmission Security

  • Transmission-security options include cryptography, physical-layer security, covert communication, quantum technology, and emerging techniques.The survey distinguishes protecting message content from concealing transmission itself.Source: B. Semantic Transmission Security · p. 18

1) Semantic Data Encryption for Secure Transmission:

  • DPJSCC separates private and public image information into distinct subcodewords, achieving near-random eavesdropping accuracy while maintaining reconstruction quality.It reduces eavesdropping accuracy by up to 18% compared with DeepJSCC and adversarial JSCC.Source: 1) Semantic Data Encryption for Secure… · p. 19

2) Physical-Layer Lightweight Secret Key Generation for SemComNet:

The survey covers key generation and several physical-layer and emerging approaches to securing semantic transmission. These methods address eavesdropping, covert transmission, and quantum-enabled security, while facing deployment constraints such as scalability and resource demands.

2) Physical-Layer Lightweight Secret Key Generation for SemComNet:

  • Physical-layer key generation derives secret keys from channel properties without key transmission, but its spatial-separation assumption may not hold in practice.The cited RIS-based scheme exploits semantic-drift randomness to enhance key generation rate.Source: 2) Physical-Layer Lightweight Secret… · pp. 19–20

3) Physical Layer Secure Transmission for SemComNet:

  • Beamforming, power allocation, cooperative interference, and artificial-noise injection widen legitimate-versus-eavesdropper channel differences, though holistic security calls for cross-layer optimization.The cited approaches include an artificial-noise-assisted interference-alignment scheme that minimizes secrecy outage probability.Source: 3) Physical Layer Secure Transmission… · p. 20

4) Covert Communication for Secure Transmission:

  • Covert communication conceals transmission from detection, while semantic steganography hides communication content; proposed schemes still face complexity and scalability concerns.A full-duplex receiver can use artificial noise to interfere with a warden's detection.Source: 4) Covert Communication for Secure… · pp. 20–21

5) Quantum Technology for Secure-enhanced SemComNet:

  • Quantum key distribution detects eavesdropping, while quantum semantic encoding protects content; quantum approaches face resource-optimization and scalability challenges.The cited quantum semantic communication work encodes data into quantum states for secure teleportation.Source: 5) Quantum Technology for… · p. 21

6) Emerging Communication Technology Assisting Secure SemComNet:

  • RIS-based semantic sensing reduces data volume by 95% without affecting sensing tasks, while passive RIS attenuation and spread-spectrum methods shape other security options.Active RIS is proposed to reconfigure wireless propagation in response to attenuation challenges.Source: 6) Emerging Communication Technology… · p. 21

C. Reliable SemComNet

Reliable SemComNet requires both resilient semantic interpretation and stable semantic-information transmission under difficult channel conditions. The survey reviews generative, adaptive, reasoning-based, and relay approaches, while noting issues such as hallucination, limited testing, and computation.

C. Reliable SemComNet

  • The reliability discussion spans accurate semantic interpretation and stable transmission under low SNR, strong perturbation, and long-distance conditions.Source: C. Reliable SemComNet · p. 21

1) GAI for Misunderstanding Mitigation in SemComNet:

  • GAI-based JSCC improves distortion and perceptual metrics over DeepJSCC even with knowledge-base mismatches, while privacy filters replace private semantic information with natural features.The cited privacy-preserving system reports improved communication efficiency and reduced misunderstanding errors.Source: 1) GAI for Misunderstanding Mitigation… · p. 21, Figure 12 · p. 22

2) Reliable Adaptive Transmission in SemComNet:

  • Adaptive transmission adjusts coding rates and uses progressive HARQ with incremental knowledge to improve efficiency, robustness, and cost-effectiveness, though extensive real-world testing is absent.The cited work evaluates results using BLEU; video bitrate adaptation is also discussed.Source: 2) Reliable Adaptive Transmission in… · p. 22

3) Reasoning-driven Reliable SemComNet:

  • Reasoning approaches include knowledge-graph-driven communication and implicit-semantic interpretation, but reported challenges include knowledge-graph typing errors and limited generalization.A signaling-game and neuro-symbolic approach is described as supporting semantic strategies with limited training samples and minimal data transmission.Source: 3) Reasoning-driven Reliable SemComNet · pp. 22–23

4) Collaborative Relay for Reliable SemComNet:

  • Semantic relay nodes forward and interpret semantic information, but the cited amplify-and-forward and decode-and-forward system faces high computational complexity and long delay.Source: 4) Collaborative Relay for Reliable… · p. 23

D. Trust Management in SemComNet

1) Authentication Mechanism in SemComNet:

The survey covers authentication, blockchain-based trust, trust evaluation, and defenses for data and knowledge security in SemComNet. It describes methods and their boundaries, including idealized channel assumptions, scalability concerns, and sparse trust information.

1) Authentication Mechanism in SemComNet:

  • Physical-layer authentication offers lightweight identity verification, but some cited schemes assume perfect channel conditions, and artificial-noise effectiveness diminishes under poor channels.Passive authentication uses channel fingerprints; active schemes embed a physical-layer tag generated from a secret key,.Source: 1) Authentication Mechanism in SemComNet · p. 23

2) Blockchain for Trustless SemComNet:

  • Blockchain supports decentralized identity, cross-domain authentication, and smart-contract authorization, while performance in large heterogeneous SemComNet remains underexplored.The cited mobile system uses self-sovereign identities and chameleon hashes for revocation.Source: 2) Blockchain for Trustless SemComNet · p. 23

3) Trust Evaluation in SemComNet:

  • Trust assessment combines direct and indirect evidence, but sparse histories and trust matrices can impair evaluations, while agent growth raises communication and computation overhead.The surveyed work also identifies difficulty acquiring experience and reputation indicators.Source: 3) Trust Evaluation in SemComNet · pp. 23–24

E. Data & Knowledge Security in SemComNet

  • Data and knowledge security is framed as a prerequisite for SemComNet, with the survey focusing on poisoning, tamperproofing, access control, and privacy preservation.Source: E. Data & Knowledge Security in… · p. 24

1) Defense Strategies Against Semantic Data Poisoning:

  • Poisoning defenses include data preprocessing, filtering, and robust training; strong augmentations are reported to improve resistance without trading off performance.Knowledge distillation can sanitize datasets by generating pseudo-labels from an auxiliary model trained on a validation set [90].Source: 1) Defense Strategies Against Semantic… · p. 24

2) Blockchain for Tamper-proof SemComNet:

  • Blockchain-based redaction and compression of blockchain content has negligible overhead, according to the cited experiments.The survey also covers provenance, integrity, and confidentiality mechanisms for semantic data and knowledge.Source: 2) Blockchain for Tamper-proof SemComNet · pp. 24–25

3) Data & Knowledge Access Control in SemComNet:

The survey reviews access-control and privacy approaches for SemComNet, spanning data and knowledge protection, machine unlearning, federated learning, and other privacy techniques. Its lessons emphasize that existing defenses offer useful mechanisms but need further development tailored to SemComNet.

3) Data & Knowledge Access Control in SemComNet:

  • Access control ranges from role- and attribute-based policies to purpose-aware models, but complexity and cross-domain access hinder adoption and context-aware control remains needed.Privacy-preserving approaches also target statistical analysis without direct access to sensitive data.Source: 3) Data & Knowledge Access Control in… · p. 25

4) Machine Unlearning for Privacy Preservation in SemComNet:

  • Federated unlearning avoids the all-training-data access required by centralized methods, with server-side and client-side approaches offering different routes to erase contributions.The cited methods include subtracting a client's accumulated updates and using knowledge distillation, or using Fisher information to reduce retraining costs.Source: 4) Machine Unlearning for Privacy… · pp. 25–26

5) FL for Privacy Preservation in SemComNet:

  • Federated learning keeps user data on devices and supports semantic-codec training and shared knowledge-base construction, including FedSem's parameter sharing and information-bottleneck mechanism.FedSem uses the information bottleneck to limit shared knowledge and improve rate-distortion and convergence performance.Source: 5) FL for Privacy Preservation in… · p. 26

6) Potential Privacy Protection Approaches for SemComNet:

  • APB-DP reduces privacy leakage rate by 13% and performance loss rate by 71% compared with the standard FL scheme, while differential-privacy noise can degrade codec accuracy.The survey also identifies knowledge distillation, secure multi-party computation, and trusted execution environments as potential privacy approaches.Source: 6) Potential Privacy Protection… · p. 26

F. Summary and Lessons Learned

  • The survey groups defenses across model security, transmission security, reliability, trust, and data and knowledge security, while identifying SemComNet-tailored privacy and security methods as an ongoing need.Its lessons include access-control policies and blockchain for data and knowledge security, alongside machine unlearning, federated learning, and differential privacy for privacy protection.Source: F. Summary and Lessons Learned · p. 26

V. FUTURE RESEARCH DIRECTIONS

The survey identifies open challenges for SemComNet research, including energy use, explainability, generative-AI integration, endogenous security, and adaptive collaboration. It concludes with a three-layer architecture, three working modes, and a review of threats, countermeasures, and future directions.

A. Green SemComNet Architecture

  • Large-scale knowledge and model sharing raises resource and energy costs, motivating green designs, infrastructure compatibility, lower-cost dataset and knowledge-base maintenance, and low-power hardware.The passage cites self-learning and self-updating datasets and knowledge bases as directions for reducing maintenance cost.Source: A. Green SemComNet Architecture · p. 27

B. Explainable Semantic Model

  • Black-box semantic codecs raise explainability concerns, while applying explainable AI across SemComNet's three layers and adapting dynamic knowledge graphs remain open challenges.The three layers are the control, semantic transmission, and cognitive sensing layers.Source: B. Explainable Semantic Model · p. 27

C. SemComNet Orchestrated with Generative AI

  • Generative AI could enrich knowledge bases and semantic reasoning, but hallucination, resource demands, and personalization remain challenges for SemComNet integration.Quantization and model compression are identified as techniques that can improve efficiency while maintaining performance.Source: C. SemComNet Orchestrated with… · p. 27

D. Endogenous Secure SemComNet

  • The survey calls for endogenous security integrated by design, but a comprehensive framework and privacy-preserving security for collaborative decisions and conflict resolution remain challenges.The envisioned approach includes self-protection, self-evolution, rapid response, and autoimmunity.Source: D. Endogenous Secure SemComNet · p. 27

E. Adaptive SemComNet Design

  • Adaptive SemComNet must address knowledge synchronization across changing topologies and incompatible task-specific semantic information, including through unified models that adapt across tasks.Continual learning, meta-learning, and multi-task learning are suggested as techniques for adaptation without extensive retraining or storing multiple models.Source: E. Adaptive SemComNet Design · pp. 27–28

VI. CONCLUSION

  • The survey presents a three-layer SemComNet architecture and discusses paired, clustered, and networked modes, enabling technologies, use cases, metrics, threats, countermeasures, and future directions.It examines security and privacy from academic and industrial perspectives.Source: §VI · p. 28
Loading 2405.01221v2…