PaperlayerView exact revision on arXiv

Source-linked AI summary

Trustworthy, Explainable, and Sustainable Decentralized Intelligence for 6G Networks

Giovanni Perin, Michele Rossi, Enrique Tomás Martínez Beltrán, Fernando Torres-Vega, José María Jorquera Valero, Manuel Gil Pérez, Eunjeong Jeong, Nikolaos Pappas, Farah Abed Zadeh, Chamara Sandeepa, Bartlomiej Siniarski, Madhusanka Liyanage, Betül Güvenç Paltun, Leyli Karaçay, Ioannis Pitsiorlas, Marios Kountouris

arXiv:2609.13872v1cs.NIcs.AIeess.SP

TL;DR

Centralized AI is restrictive for 6G security because timely decisions, heterogeneous data, and multi-domain constraints challenge centralized pipelines. The paper presents decentralized learning alongside trust, explainability, and sustainable-computing considerations; it concludes that these dimensions require joint design, while realistic deployments still face substantial operational challenges.

  • Problem

    Centralized 6G AI faces data-volume, latency, and multi-domain governance constraints, including security tasks that need decisions close to the point of observation.

    Source: Limitations of Centralized AI for 6G · p. 3

  • Method

    The paper develops a unified perspective on decentralized 6G intelligence, including graph-regularized consensus and joint energy- and age-of-information-aware training schedulers.

    Source: Distributed Model Training: Federated… · p. 5, CONCLUDING REMARKS · p. 18

  • Results

    The paper concludes that decentralization, trustworthiness, explainability, and sustainability should jointly shape the 6G security stack rather than be treated as separate mechanisms.

    Source: CONCLUDING REMARKS · p. 18

  • Takeaways & Limitations

    The proposed direction combines low-power neuromorphic computing with joint energy- and age-of-information-aware training schedulers to keep security mechanisms computationally viable without excessive resource drain.

    Source: CONCLUDING REMARKS · p. 18

    Realistic 6G FL and DFL must contend with heterogeneous capacity, non-IID data, changing channels, intermittent participation, adaptive adversaries, and privacy or security overhead.

    Source: Distributed Intelligence and… · p. 7

Abstract

from arXiv · show

As 6G networks transition from theoretical frameworks into operational realities, artificial intelligence (AI) evolves from an add-on optimization tool into a distributed and interconnected structural layer. Unlike previous network generations that mostly relied on centralized cloud analytics platforms, AI-native 6G networks operate across a dynamic, multi-domain edge-cloud continuum where data originates from heterogeneous sources including user devices, radio access networks, sensing infrastructures, and vertical applications. Centralizing this massive volume of data creates severe communication overhead, unacceptable latency bottlenecks, single points of failure, and complex cross-domain governance challenges. Consequently, decentralization becomes a fundamental architectural requirement for future 6G network intelligence and zero-touch operations. Security serves as the primary enabler of this decentralized paradigm. Critical security functions, such as real-time threat detection, physical-layer attack mitigation, slice protection, and intrusion detection, require immediate access to local context and telemetry before operational data loses its value. However, moving intelligence to the edge via collaborative paradigms like federated learning (FL) and decentralized FL (DFL) introduces complex trade-offs. System security cannot be addressed in isolation; it is deeply intertwined with equally important aspects like trustworthiness, explainability, and energy sustainability. Taking these aspects into account, this paper develops a unified perspective on decentralized intelligence for 6G, arguing that decentralization, trustworthiness, explainability, and sustainability must be designed jointly rather than treated as independent requirements

DECENTRALIZED INTELLIGENCE IN AI-NATIVE 6G SYSTEMS

Centralized AI becomes restrictive for 6G because data volume, real-time security needs, and multi-domain governance strain centralized pipelines. FL keeps raw data local, while DFL replaces the central aggregator with peer-to-peer collaboration over potentially changing topologies.

Limitations of Centralized AI for 6G

  • Centralized collection of heterogeneous 6G data can create communication overhead and energy consumption, while central analysis may miss real-time security control timescales.Multi-domain stakeholders may also retain distinct privacy, regulatory, and business constraints, making centralized ownership often unrealistic.Source: Limitations of Centralized AI for 6G · p. 3

Distributed Model Training: Federated and Decentralized Learning

  • FL supports collaborative security-model training by keeping each participant’s observations within its own domain.The passage names intrusion detection, anomaly detection, traffic classification, RF fingerprinting, and attack prediction as potential uses.Source: Distributed Model Training: Federated… · p. 4
  • DFL replaces FL’s central aggregator with peer-to-peer model exchange, local optimization, neighborhood aggregation, and progressive agreement among peers.Conventional FL’s parameter server can be a communication bottleneck, privacy-sensitive entity, attack target, and single point of failure.Source: Distributed Model Training: Federated… · p. 4
  • A weighted communication graph represents participating nodes and active links, while graph-regularized DFL encourages neighboring models to agree.The local objectives are weighted by p_k, and λ controls the overall strength of consensus regularization.Source: Distributed Model Training: Federated… · pp. 4–5
  • DFL can follow horizontal or vertical 6G infrastructure relationships and adapt collaboration to heterogeneous, intermittent, multi-domain participation.Updates may be weighted, filtered, or discarded according to estimated trustworthiness, informational value, and cost.Source: Distributed Model Training: Federated… · p. 6

Distributed Intelligence and Zero-Touch Management

Decentralized intelligence can join learning and control in a zero-touch feedback loop across the edge–cloud continuum. Realizing this vision under 6G conditions requires jointly addressing operational constraints and trustworthiness requirements.

Distributed Intelligence and Zero-Touch Management

  • Local observations update distributed models, whose outputs support network decisions; resulting actions change network state and may prompt further adaptation.Edge and far-edge nodes can provide model updates, anomaly scores, confidence indicators, resource constraints, and trust-related metadata to orchestration functions.Source: Distributed Intelligence and… · pp. 6–7
  • Before distributed outputs enter automated control loops, learning processes and models must be assessed for robustness, privacy, fairness, explainability, accountability, sustainability, and node or update reliability.Source: Distributed Intelligence and… · p. 7
  • Realistic 6G conditions include heterogeneous node capacity, non-IID data, variable channels, intermittent participation, adaptive adversaries, and overhead from privacy- and security-preserving mechanisms.Future FL and DFL systems must jointly address accuracy, communication cost, energy use, privacy leakage, robustness, and operational transparency.Source: Distributed Intelligence and… · p. 7

Robustness to Manipulated Input

6G AI systems face attacks during training and inference, with manipulated inputs or updates able to produce incorrect models or predictions. The paper describes defenses spanning data validation, distributed-update assessment, robust aggregation, and model-level resilience.

Robustness to Manipulated Input

  • Poisoning can manipulate training data, labels, gradients, or updates, while adversarial inputs can cause incorrect predictions with potential network consequences.Examples include incorrect resource allocation, degraded quality of service, or large-scale network disruptions.Source: Robustness to Manipulated Input · p. 7
  • A holistic defense combines data validation and sanitization, anomaly detection, reputation-based filtering, and robust aggregation such as median-based or trimmed-mean methods.The passage also identifies adversarial training,, uncertainty estimation, and ensemble learning as model-level measures.Source: Robustness to Manipulated Input · p. 8

Ensuring Data Privacy

Privacy-preserving AI seeks to protect data and model confidentiality across the AI lifecycle, but its mechanisms trade off against accuracy, latency, resources, and scalability. In 6G security, explainability can also support intrusion-detection operations, provided explanations and prediction reliability are treated as distinct concerns.

Ensuring Data Privacy

  • Privacy-preserving AI limits sensitive-data exposure throughout the AI lifecycle, while model privacy also concerns information inferred from outputs, parameters, gradients, updates, or interfaces [8].Potential inferences include training-record membership, private-feature reconstruction, and model extraction.Source: Ensuring Data Privacy · p. 8
  • Federated learning keeps data localized, while secure aggregation and methods such as HE, SMC, TEEs, and differential privacy provide complementary protections.These mechanisms can increase computational overhead or create trade-offs in accuracy, latency, resource use, and scalability.Source: Ensuring Data Privacy · pp. 8–9
  • Peer-to-peer DFL can encrypt local updates and use collaborative secure aggregation without a central aggregator, but requires coordination and key-management techniques.Model updates can still leak sensitive information, and validation or attack detection without plaintext updates remains challenging.Source: Ensuring Data Privacy · p. 9

Explainability, Transparency, and Trust in 6G AI

  • For intrusion detection, explanations can help catch model errors and support trust in automated security, but an explanation does not establish that a decision is dependable enough for action.The paper calls for explainability across the IDS lifecycle and notes that explanation methods themselves can be manipulated.Source: Explainability, Transparency, and… · p. 10
  • IDS explanations and attributions can support model optimization, trust assessment, and automated security operations.The paper identifies SHAP and LIME as examples of explanation methods and warns that misleading explanations can conceal malicious behavior.Source: Explainability, Transparency, and… · pp. 10–11

Explanations as Actionable, Trustworthy Signals

The section treats explanations as operational signals for IDS lifecycle management and security actions, but argues they need reliability estimates and trustworthy, robust handling before automated use.

Explanations as Actionable, Trustworthy Signals

  • Prediction explanations do not establish reliability, so automated IDS actions need accompanying confidence or uncertainty estimates.Uncertain cases near decision boundaries or in sparsely represented regions may require validation or human escalation.Source: Explanations as Actionable,… · p. 11
  • Attributions can guide feature reduction and ANN pruning, potentially shrinking models, reducing attacker-manipulable inputs, and meeting edge memory, latency, and energy budgets.Tracking attribution distributions can also reveal changing model reliance and prompt retraining.Source: Explanations as Actionable,… · pp. 11–12
  • Attributions can support fine-grained automated responses and let peers in decentralized detection check alerts against their own models and policies.Schema-conformant records can guide targeted mitigation, access decisions, or human escalation.Source: Explanations as Actionable,… · p. 12
  • Post-hoc attribution methods are not adversarially robust by construction, so explanations used for policy must be robust, privacy-controlled, and faithful.Manipulated explanations can leave predictions unchanged while presenting arbitrary or benign-looking rationales.Source: Explanations as Actionable,… · p. 12
  • Attribution shifts may flag concept drift before accuracy degrades, including shifts that leave input statistics unchanged while changing the learned input–output relationship.The section presents this as a hypothesis for investigation, not an established result.Source: Explanations as Actionable,… · pp. 12–13
  • Tracking attribution distributions over traffic windows may expose attacks on the explainer, once trustworthy baselines and realistic tamper models are formalized.This is the second proposed hypothesis.Source: Explanations as Actionable,… · p. 13

SUSTAINABLE INTELLIGENCE AT THE EDGE

The section treats sustainability as a design constraint for edge intelligence and contrasts limited post-training efficiency gains with energy-efficient-by-design models and sustainable computing infrastructure.

SUSTAINABLE INTELLIGENCE AT THE EDGE

  • Sustainability must constrain edge and far-edge learning, inference, and security mechanisms within their energy budgets.The section argues energy efficiency should be pursued from the network design phase, not only optimized after architecture choices.Source: SUSTAINABLE INTELLIGENCE AT THE EDGE · p. 13
  • The section identifies two directions: energy-efficient AI/ML models on dedicated ultra-low-power hardware, and sustainable network computing through renewables, storage management, and computation scheduling.These directions decentralize in-network processing.Source: SUSTAINABLE INTELLIGENCE AT THE EDGE · p. 13

Natively Energy Efficient AI Models

  • Pruning removes low-magnitude parameters followed by retraining, while quantization reduces the bits used to store parameters after training.Both are presented as context-agnostic post-training efficiency methods.Source: Natively Energy Efficient AI Models · p. 14
  • Post-training pruning and quantization can reduce memory and computation with limited accuracy degradation, but their efficiency improvements are limited.The passage reports that 8-bit floating-point use does not significantly degrade accuracy, while improving memory footprint and computational power.Source: Natively Energy Efficient AI Models · p. 14
  • The section advocates energy-efficient-by-design deep learning, identifying reservoir computing and spiking neural networks as suitable neuromorphic paradigms.This contrasts with relying only on post-training operations.Source: Natively Energy Efficient AI Models · p. 14

Reservoir computing:

Reservoir computing uses fixed recurrent dynamics with a trained output layer to process temporal signals, while spiking neural networks encode signals as discrete spikes and update stateful neuron potentials over time.

Reservoir computing:

  • Reservoir computing projects inputs into a high-dimensional temporal representation using fixed recurrent dynamics, training only a shallow output layer.The reduced training burden lowers computational complexity and energy consumption, and the retained temporal information suits sequential signals.Source: Reservoir computing · p. 14
  • Liquid state machines connect reservoir computing to spiking networks because their reservoirs consist of recurrently connected spiking neurons.Echo state networks are another cited reservoir-computing approach.Source: Reservoir computing · p. 14

Spiking neural networks:

  • Spiking neural networks propagate signals through emitted current spikes encoded as binary values: 1 denotes spike emission, while 0 means no signal propagation.Source: Spiking neural networks · p. 14
  • A leaky integrate-and-fire neuron models an RC circuit and updates its membrane potential as a state that charges with input current and exponentially discharges over time.When the potential reaches a firing threshold, the neuron emits a spike and resets to a reset threshold; τm = RC sets the discharge timescale.Source: Spiking neural networks · p. 15

Hardware Implementations:

The section describes dedicated electronic and photonic implementations of neuromorphic computing and reports energy-delay-product gains for SNNs on dedicated hardware.

Hardware Implementations:

  • SNN hardware can use dedicated digital, analog, or mixed-signal circuits, including CMOS and emerging memristive technologies.These provide different realizations of stateful neuronal and synaptic computation.Source: Hardware Implementations · p. 15
  • Photonic reservoir computing uses optical systems’ fast dynamics and low propagation losses for high-throughput, energy-efficient temporal-signal processing.Source: Hardware Implementations · p. 15
  • Up to three orders of magnitude improvement in energy-delay-product was reported for SNNs running on an Intel Loihi processor.The cited study presents dedicated SNN hardware as a promising candidate to break the traditional energy-versus-latency tradeoff.Source: Hardware Implementations · p. 16

Energy- and Context-Aware Distributed Learning

Sustainable distributed learning must account for both the energy cost of local computation and the information value of each update. Scheduling should jointly consider device energy and update context, especially under scarce energy and heterogeneous data.

Energy- and Context-Aware Distributed Learning

  • For energy-harvesting devices, redundant local training uses energy that could otherwise support a more informative future update, so scheduling should consider both energy state and operational context.Battery availability alone may schedule updates that contribute little to the global model.Source: Energy- and Context-Aware Distributed… · p. 16
  • Group-based turn-taking defers local training until shortly before transmission, avoiding premature training on stale global models and smoothing active participation across rounds.The scheduler must bound the controlled staleness this introduces.Source: Energy- and Context-Aware Distributed… · p. 16
  • Age-of-information variants can favor fresh, substantive updates, while lightweight comparisons of compact intermediate representations offer a less costly proxy than direct parameter-space divergence [12].Energy awareness alone cannot assess update value; information awareness alone cannot ensure that an update is feasible to produce.Source: Energy- and Context-Aware Distributed… · pp. 16–17
  • Joint energy- and information-aware scheduling addresses the complementary limits of conserving resources and selecting valuable updates, with benefits most pronounced under severe data heterogeneity and scarce energy.The approach can also extend from deciding whether a device trains to deciding how much it trains.Source: Energy- and Context-Aware Distributed… · p. 17

CONCLUDING REMARKS

The conclusion calls for 6G security intelligence to jointly address decentralization, trust, explainability, and energy constraints. It pairs actionable explainability signals with low-power neuromorphic computing and energy- and age-of-information-aware scheduling.

CONCLUDING REMARKS

  • A unified design strategy moves intelligence to peer-to-peer decentralized architectures while treating trust as a core requirement for protecting training, algorithmic trustworthiness, and data confidentiality.The paper argues that decentralization alone is insufficient if the underlying intelligence lacks trust.Source: CONCLUDING REMARKS · p. 17
  • Explainability tools should provide operational signals that quantify uncertainty, detect concept drift, and validate alerts across peers for automated decision-making.Source: CONCLUDING REMARKS · p. 17
  • The paper proposes low-power neuromorphic architectures and joint energy- and age-of-information-aware training schedulers to keep distributed security mechanisms computationally viable within strict energy budgets.Source: CONCLUDING REMARKS · p. 18
  • Decentralization, security, explainability, and sustainability should jointly shape learning participation, update trust, model adaptation, decision reliability, and energy expenditure.The paper rejects treating these dimensions as separate mechanisms appended to decentralized AI.Source: CONCLUDING REMARKS · p. 18

BIOGRAPHIES

The biographies describe contributors working across distributed learning, network security, trustworthy and explainable AI, and sustainable networking. Their listed research interests span related areas including federated learning, cybersecurity, and energy-efficient scheduling.

BIOGRAPHIES

  • Listed research interests include distributed learning, optimization, sustainable AI, networks, and energy-efficient scheduling.Source: BIOGRAPHIES · p. 19
  • Several biographies identify interests in AI security, network security, cybersecurity, intrusion detection, and trust and reputation management.Source: BIOGRAPHIES · pp. 19–20
  • His research interests include 5G/6G security, artificial intelligence, explainable AI, federated learning, blockchain, and edge computing; he is also identified as a Senior Member of IEEE.Source: BIOGRAPHIES · p. 20
Loading 2609.13872v1…