PaperlayerView exact revision on arXiv

Source-linked AI summary

Cybersecurity in Edge Computing: A Trust-Aware Federated Hybrid Intrusion Detection Framework

Zawad Yalmie Sazid, Robert Abbas

arXiv:2609.39584v1cs.CRcs.LG

TL;DR

Edge systems face a trade-off between decentralized processing and protecting telemetry across vulnerable, resource-constrained environments. TA-FHIDF combines localized hybrid detection with federated training and trust-aware aggregation; the paper reports stable global detection under sustained adversarial attack.

Abstract

from arXiv · show

Edge computing has emerged as a critical computing paradigm in modern distributed systems by migrating data processing closer to end users and Internet of Things (IoT) devices. While this paradigm decentralizes processes, minimizes latency, and reduces backhaul bandwidth congestion, it exponentially enlarges the cyberattack surface. Heterogeneous, resource-constrained edge devices deployed across unmanaged administrative domains present highly vulnerable targets. To address these vulnerabilities without compromising global data privacy regulations, this paper proposes a novel Trust-Aware Federated Hybrid Intrusion Detection Framework (TA-FHIDF). The proposed framework integrates an Autoencoder, a 1D Convolutional Neural Network (1D-CNN), and a Bidirectional Long Short-Term Memory (BiLSTM) model into a unified, localized deep learning engine capable of autonomous spatial and temporal feature extraction. Model training is performed collaboratively via federated learning, ensuring raw network telemetry remains isolated at local gateways. Furthermore, to defend against adversarial model poisoning attacks, we introduce a robust server-side trust-aware aggregation mechanism that evaluates client reliability using a cosine similarity metric before global model integration. Empirical evaluations across multi-vector benchmark datasets (UNSW-NB15, CICIDS2017, and Edge-IIoTset) demonstrate the framework's superior detection accuracy, rapid convergence, and high Byzantine fault tolerance under adversarial attack scenarios.

I. INTRODUCTION

Edge computing brings processing closer to users and IoT devices, but expands the attack surface across vulnerable, resource-constrained edge environments. The paper frames the challenge as balancing privacy, detection accuracy, poisoning resistance, and hardware feasibility, then sets objectives spanning hybrid detection, federated learning, trust-aware aggregation, and comparative evaluation.

A. Research Background

  • Edge decentralization expands the attack surface, while heterogeneous, resource-constrained devices in unmanaged domains are susceptible to localized exploitation,.The paper also notes that edge systems must preserve privacy, minimize latency, and meet regulatory requirements.Source: A. Research Background · p. 1, B. Existing Challenges · p. 1

B. Existing Challenges

C. Research Problem and Objectives

  • The objectives are a localized Autoencoder–1D-CNN–BiLSTM detector, federated training without raw-telemetry backhaul, cosine-similarity update filtering, and comparative ablation against traditional classifiers.The planned evaluation includes detection accuracy and false-positive reduction across multiple benchmark datasets.Source: C. Research Problem and Objectives · p. 2

II. LITERATURE REVIEW

The literature review traces gaps across edge security, hybrid intrusion detection, and privacy-preserving collaborative learning. It motivates a federated hybrid detector that combines edge-local processing with defenses against untrusted client updates.

A. Foundations of Edge Computing and Network Security

  • Prior edge-computing work established latency and bandwidth motivations, while security concerns include vulnerable edge nodes and centralized systems’ latency costs; blockchain consensus overhead and static datasets are also cited as limitations,,,,.The review identifies a need for a robust decentralized defense at the physical edge layer.Source: A. Foundations of Edge Computing and… · p. 2

B. Deep Learning and Hybrid Architectures for IDS

  • Prior IDS approaches include manual-feature edge classifiers and attack-specific models, while hybrid deep models face memory, privacy, or collaborative-learning gaps,,,,.The review positions AE-CNN-BiLSTM as combining spatial compression with federated training at the edge.Source: B. Deep Learning and Hybrid… · p. 2

C. Privacy-Preserving Collaborative Learning

  • FedAvg avoids raw-telemetry transmission, but the review notes its trust assumption and sensitivity to non-IID gradients; incremental federated learning is also described as high-latency and lacking server-side cryptographic validation,,.The review motivates robust aggregation and cites Edge-IIoTset as a benchmark tailored to distributed learning.Source: C. Privacy-Preserving Collaborative… · p. 2

D. Trust-Aware and Resilient Aggregation Mechanisms

The review identifies a gap in integrating mathematical trust evaluation with hybrid deep models for multi-vector attacks at physical gateways. The methodology pairs constrained-edge feature processing with an Autoencoder, 1D-CNN, and BiLSTM pipeline.

D. Trust-Aware and Resilient Aggregation Mechanisms

  • Existing trust approaches are described as limited by heuristic reliability signals, identity-based defenses that can miss compromised authenticated devices, or unverified scalability to deep models,,.The review identifies limited integration of deep cosine-similarity inspection with AE-CNN-BiLSTM as the remaining gap, which the proposed framework aims to address [3],.Source: D. Trust-Aware and Resilient… · p. 3

III. METHODOLOGY AND SYSTEM ARCHITECTURE

  • TA-FHIDF is designed for strict hardware constraints, data privacy, and defense against malicious federation updates, with threat detection decentralized across three interconnected layers.Its local edge-node design places a multi-stage feature-extraction topology at the gateway.Source: §III · p. 3, B. Local Edge Node Architecture… · p. 3

A. Data Ingestion and Preprocessing Pipeline

  • Local preprocessing removes duplicate flows and zero-variance columns, median-imputes missing features, one-hot encodes categorical features, scales continuous values to [0, 1], and windows data into sequential tensors.The input telemetry is characterized as noisy, high-dimensional, and statistically skewed.Source: A. Data Ingestion and Preprocessing… · p. 3

B. Local Edge Node Architecture (AE-CNN-BiLSTM)

  • The Autoencoder compresses high-dimensional input into latent features for constrained hardware, and the 1D-CNN extracts localized spatial correlations from those features.The CNN’s kernel size is k; the Autoencoder’s encoder uses weight matrix We, bias be, and ReLU activation σ.Source: B. Local Edge Node Architecture… · p. 3
  • The CNN outputs feed a BiLSTM that captures multi-stage attack sequences through forward and backward hidden states before Softmax classification.The two directional states form the temporal representation.Source: B. Local Edge Node Architecture… · p. 3

C. Privacy-Preserving Federated Training

Federated training keeps raw network telemetry at local edge nodes while exchanging secured model updates. A server-side trust mechanism scores client updates before aggregation, and the paper reports that this intervention isolates and throttles malicious or heavily non-IID updates.

C. Privacy-Preserving Federated Training

  • Edge nodes train on isolated datasets, and only cryptographically secured parameter updates—not raw telemetry—cross administrative boundaries.The paper states this approach eliminates backhaul congestion and adheres to privacy mandates.Source: C. Privacy-Preserving Federated Training · p. 3

D. Server-Side Trust-Aware Aggregation

  • Before aggregation, the server compares each client update with the global average update using cosine similarity to calculate a dynamic trust score, bypassing standard FedAvg.The paper describes this mechanism as a response to adversarial compromise, including label flipping and gradient corruption.Source: D. Server-Side Trust-Aware Aggregation · p. 3
  • The global model uses a trust-score-weighted average, and the paper states that this intervention isolates and throttles malicious or heavily non-IID updates to support Byzantine resilience and stable convergence.The passage presents these as algorithmic effects of the trust-aware intervention.Source: D. Server-Side Trust-Aware Aggregation · pp. 3–4

E. Proposed Architectural Framework

The framework addresses edge hardware limits, privacy regulations, and compromised-node threats through a unified hybrid intrusion-detection architecture with trust-aware decentralized learning. Its localized engine automates feature extraction, while local training keeps raw telemetry within administrative boundaries.

E. Proposed Architectural Framework

  • The framework combines hybrid intrusion detection with trust-aware decentralized teamwork to address edge resource limits, privacy requirements, and compromised-node threats.The architecture also performs real-time threat detection at the edge without sending raw network traffic to a central cloud.Source: E. Proposed Architectural Framework · p. 4

1) Architecture Overview and Data Ingestion:

  • The architecture begins with heterogeneous IoT and IIoT traffic, which localized edge nodes preprocess for deep-learning ingestion.The supplied passages do not specify the preprocessing steps in this subsection.Source: 1) Architecture Overview and Data… · p. 4

2) Unified Hybrid Deep Learning Engine:

  • A tripartite detection pipeline runs at the edge gateway and automates feature extraction without manual engineering.The passage identifies the pipeline as the core detection engine but does not name its three components here.Source: 2) Unified Hybrid Deep Learning Engine · p. 4

3) Decentralized Federated Learning Pipeline:

  • Edge nodes train locally on isolated datasets, keeping raw network telemetry within local administrative boundaries during collaborative learning.The central server evaluates client updates using cosine similarity; the passage also describes blocking malicious traffic and quarantining compromised endpoints after anomaly detection.Source: 3) Decentralized Federated Learning… · p. 4

F. Dataflow Diagram for the Model

The dataflow diagram follows telemetry through local preprocessing and detection, then sends client updates to the cloud for trust evaluation and aggregation. The resulting global model returns to edge gateways for the next training round.

Data Ingestion and Local Storage

  • At the edge, IoT telemetry is cleaned, scaled, and reshaped before storage as preprocessed sequential tensors in the local edge dataset.Training batches are drawn from that local store for the hybrid engine, which sends intrusion classifications to the firewall or network controller for mitigation.Source: F. Dataflow Diagram for the Model · p. 4

Trust Evaluation and Global Consensus (The Cloud Loop)

  • Only localized mathematical gradients, sent as the Local Parameter Update ∆w, cross the external network boundary, preserving local data privacy.The framework evaluates transmitted gradients for statistical anomalies or Byzantine poisoning before aggregation.Source: F. Dataflow Diagram for the Model · p. 4
  • The cloud evaluates updates using cosine similarity, then passes a Validated Update and Trust Score T to secure federated aggregation.Aggregation filters heavily penalized parameters and computes a weighted average.Source: F. Dataflow Diagram for the Model · p. 4
  • The aggregated New Global Model wt+1 is stored in the Global Model Registry and broadcast back to edge gateways to begin another collaborative-training round.Source: F. Dataflow Diagram for the Model · p. 4

IV. DATA COLLECTION AND PROCESSING

The evaluation pipeline uses three cybersecurity benchmarks and describes processing steps intended to prepare heterogeneous traffic for deep learning and federated edge simulation. The dataset choices span general intrusion scenarios, complex contemporary attacks, and IoT/IIoT settings.

A. Datasets and Feature Statistics

  • The evaluation uses UNSW-NB15, CICIDS2017, and Edge-IIoTset to test the framework across diverse threat vectors and network settings.The multi-dataset strategy is intended to support generalizability and avoid reliance on a single network topology.Source: A. Datasets and Feature Statistics · p. 5
  • UNSW-NB15 supplies 49 flow features and nine attack families, while CICIDS2017 provides over 80 flow-based features and multi-day attack scenarios.The passages describe UNSW-NB15 as a baseline benchmark and CICIDS2017 as useful for testing spatial extraction and dimensionality reduction.Source: A. Datasets and Feature Statistics · p. 5
  • Edge-IIoTset covers physical IoT devices and lightweight protocols, providing a testbed for federated learning, hardware constraints, localized traffic, and data heterogeneity.Source: A. Datasets and Feature Statistics · p. 5

B. Data Processing Criteria and Strategies

  • The pipeline removes duplicate rows and zero-variance features to reduce edge-device computation and prevent biased heuristics.It also uses median substitution for missing values because network telemetry may contain extreme outliers.Source: B. Data Processing Criteria and… · p. 6
  • Categorical protocol and service features are converted to numerical inputs, and continuous features are min-max scaled to [0, 1].The passage describes one-hot encoding as the likely categorical encoding approach.Source: B. Data Processing Criteria and… · pp. 6–7
  • Normalized 2D traffic data is reshaped into 3D sequential tensors so the 1D-CNN and BiLSTM can process temporal sequences.An overlapping sliding window lets the BiLSTM capture prolonged, multi-stage attack behaviors.Source: B. Data Processing Criteria and… · p. 7
  • Dirichlet-distributed partitions create non-IID client data to stress-test Byzantine resilience and global stability under heterogeneous edge conditions.The concentration parameter α governs the simulated class imbalance.Source: B. Data Processing Criteria and… · p. 7

V. EXPERIMENTAL EVALUATION AND RESULTS ANALYSIS

The section describes an empirical evaluation using three standardized cybersecurity benchmarks. The multi-dataset design is intended to assess generalizability and diverse threat vectors across four research objectives.

V. EXPERIMENTAL EVALUATION AND RESULTS ANALYSIS

  • The framework is evaluated using a triad of modern, standardized cybersecurity benchmark datasets.Source: §V · p. 7
  • The multi-dataset strategy is intended to test generalizability and performance across diverse threat vectors.Source: §V · p. 7
  • The evaluation is organized around four primary research objectives.Source: §V · p. 7

A. Evaluation of the Centralized Hybrid Deep Learning Engine

The centralized AE-CNN-BiLSTM engine achieved high CICIDS2017 DDoS classification performance; federated evaluation reached 99% global accuracy by round two and reported strong F1-scores across attack vectors.

A. Evaluation of the Centralized Hybrid Deep Learning Engine

B. Evaluation of the Privacy-Preserving Collaborative Framework

C. Evaluation of the Trust-Aware Byzantine-Resilient Aggregation

Under simulated poisoned updates, global accuracy stabilized near 94%; cosine-similarity trust evaluation depressed the malicious gateway's score while maintaining stable scores for benign gateways.

C. Evaluation of the Trust-Aware Byzantine-Resilient Aggregation

D. Comparative Benchmarking and Ablation Study

The study compares traditional classifiers on UNSW-NB15 and concludes that the framework maintained a 94% global detection rate under sustained adversarial attack. It also identifies future extensions for concept drift, latency variation, and energy use.

D. Comparative Benchmarking and Ablation Study

VI. CONCLUSION AND FUTURE WORK

  • The localized Autoencoder, 1D-CNN, and BiLSTM engine extracts spatial and temporal threat features; federated learning keeps raw telemetry local and eliminates backhaul congestion.Source: §VI · p. 9
  • 94% global detection rate was maintained under sustained adversarial attack, with poisoned client updates isolated by the trust metric.The conclusion attributes this Byzantine fault tolerance to a dynamic, server-side cosine-distance trust metric.Source: §VI · p. 9
  • Future work proposes incremental federated learning for concept drift, adaptive optimization for latency fluctuations, and energy-aware processing for battery-constrained sensors.Source: §VI · p. 9
Loading 2609.39584v1…