Paper page - EdgeDetect: Importance-Aware Gradient Compression with Homomorphic Aggregation for Federated Intrusion Detection

Source: https://huggingface.co/papers/2604.14663

https://huggingface.co/papers/2604.14663#edgedetect-importance-aware-gradient-compression-with-homomorphic-aggregation-for-federated-intrusion-detectionEdgeDetect: Importance-Aware Gradient Compression with Homomorphic Aggregation for Federated Intrusion Detection

**Authors:**Noor Islam S. Mohammad **Affiliation:**Department of Computer Science, Istanbul Technical University, Maslak, TR Email:[email protected] **arXiv:**2604.14663v1 [cs.CR] **Date:**16 Apr 2026

https://huggingface.co/papers/2604.14663#abstractAbstract

Federated learning (FL) enables collaborative intrusion detection without raw data exchange, but conventional FL incurs high communication overhead from full-precision gradient transmission and remains vulnerable to gradient inference attacks. This paper presentsEdgeDetect, a communication-efficient and privacy-aware federated IDS for bandwidth-constrained 6G-IoT environments.

EdgeDetect introducesgradient smartification, a median-based statistical binarization that compresses local updates to {+1, −1} representations, reducing uplink payload by**32×**while preserving convergence. We further integrate Paillier homomorphic encryption over binarized gradients, protecting against honest-but-curious servers without exposing individual updates.

Key Results:

• 98.0% multi-class accuracyand97.9% macro F1-scoreon CIC-IDS2017 (2.8M flows, 7 attack classes)
• 96.9% communication reduction(450 MB → 14 MB per round)
• Raspberry Pi 4 deployment: 4.2 MB memory, 0.8 ms latency, 12 mJ per inference with <0.5% accuracy loss
• Robustness: Maintains 87% accuracy under 5% poisoning attacks with 0.95 minority class F1 (p < 0.001)

**Index Terms:**PPFL, IDS, Edge Computing, 6G Security, IoT Networks, Communication Efficiency, Machine Learning

https://huggingface.co/papers/2604.14663#1-introduction1. Introduction

Next-generation wireless technologies (5G, 6G, IoT) enable massive machine-type communications while expanding attack surfaces for sophisticated cyber threats. Traditional centralized IDS face:

• Scalability bottlenecks
• Communication latency
• Single points of failure
• Difficulty handling high dimensionality and severe class imbalance

Federated Learningaddresses this but faces two critical challenges:

1. Communication overhead- High-dimensional gradient vectors consume excessive bandwidth
2. Gradient leakage- Shared updates may be reverse-engineered to reconstruct sensitive training samples

https://huggingface.co/papers/2604.14663#contributionsContributions

https://huggingface.co/papers/2604.14663#1-alignment-aware-federated-ids-architecture1.Alignment-Aware Federated IDS Architecture

• Privacy-preserving federated intrusion detection framework for 6G-IoT
• Integrates PCA-based dimensionality reduction, imbalance-aware sampling, and secure aggregation
• Enables collaborative learning without sharing raw network traffic

https://huggingface.co/papers/2604.14663#2-adaptive-median-based-gradient-smartification-with-encrypted-aggregation2.Adaptive Median-Based Gradient Smartification with Encrypted Aggregation

• Statistically adaptive median-threshold binarization strategy
• Compresses gradients into {+1, −1} while preserving directional alignment
• Combined with Paillier homomorphic encryption
• Achievesup to 32× communication reductionwhile mitigating gradient inversion risks

https://huggingface.co/papers/2604.14663#3-quantified-privacyutilityefficiency-trade-off3.Quantified Privacy–Utility–Efficiency Trade-off

• Extensive ablation and adversarial analyses
• 98.0% multi-class accuracy with 96.9% communication reduction
• Performance comparable to centralized baselines
• Cryptographic privacy guarantees
• Maintains >85% accuracy with 20% malicious clients
• Reduces inversion PSNR from 31.7 dB to 15.1 dB

https://huggingface.co/papers/2604.14663#4-edge-validated-deployment4.Edge-Validated Deployment

• Real-world deployment on Raspberry Pi 4
• Only 4.2 MB memory, 0.8 ms latency
• 12 mJ per inference with <0.5% accuracy degradation
• Validates suitability for resource-constrained 6G-IoT environments

https://huggingface.co/papers/2604.14663#2-related-work2. Related Work

https://huggingface.co/papers/2604.14663#a-deep-learning-based-anomaly-detectionA. Deep Learning-Based Anomaly Detection

• CNN–RNN and LSTM architectures for DDoS and zero-day detection
• Image-based encodings of time-series traffic for spatial feature extraction
• SVMs and random forests remain competitive for structured features

https://huggingface.co/papers/2604.14663#b-federated-learning-in-iot-networksB. Federated Learning in IoT Networks

• Enables decentralized training without sharing raw data
• Applications: IoT security, industrial sensor networks, cross-domain intrusion detection
• Edge–cloud collaborative architectures reduce response latency
• Challenge: Standard FL (FedAvg) relies on full-precision gradient exchange

https://huggingface.co/papers/2604.14663#c-privacy-preservation-and-gradient-compressionC. Privacy Preservation and Gradient Compression

• **Differential Privacy (DP)andHomomorphic Encryption (HE)**improve confidentiality
• Communication-efficient methods: signSGD, gradient sparsification
• Few approaches jointly optimize gradient compression and encrypted aggregation

https://huggingface.co/papers/2604.14663#d-distinction-from-signsgd-and-quantized-flD. Distinction from signSGD and Quantized FL

Unlike fixed-threshold quantizers (QSGD, TernGrad):

• Adaptive per-client thresholdadapts to gradient distribution
• Preserves relative orderingwithin each gradient vector
• Exploits heavy-tailed distributionstypical in IDS models

https://huggingface.co/papers/2604.14663#3-system-architecture3. System Architecture

https://huggingface.co/papers/2604.14663#protocol-flowProtocol Flow

EdgeDetect comprises K resource-constrained edge clients and a central aggregation server.

https://huggingface.co/papers/2604.14663#phase-1-client-side-local-trainingPhase 1: Client-Side Local Training

W_{i}^{(r+1)} = W_{i}^{(r)} − η∇L(W_{i}^{(r)}, D_i)
Δ_{i}^{(r)} = W_{i}^{(r+1)} − W^{(r)}

https://huggingface.co/papers/2604.14663#phase-2-gradient-smartificationPhase 2: Gradient Smartification

θ_i = median(|Δ_{i}^{(r)}|)
Δ^{bin}_{i,j} = +1  if Δ_{i,j} ≥ θ_i
                -1  otherwise

https://huggingface.co/papers/2604.14663#phase-3-privacy-preserving-encryptionPhase 3: Privacy-Preserving Encryption

C_{i}^{(r)} = E(Δ^{bin}_{i})  # Paillier encryption

https://huggingface.co/papers/2604.14663#phase-4-secure-aggregation-and-global-updatePhase 4: Secure Aggregation and Global Update

Δ^{bin}_{agg} = (1/|S_r|) × Σ D(C_{i}^{(r)})
W^{(r+1)} = W^{(r)} + α · Δ^{bin}_{agg}

https://huggingface.co/papers/2604.14663#4-methodology4. Methodology

https://huggingface.co/papers/2604.14663#a-data-exploration-and-preprocessingA. Data Exploration and Preprocessing

CIC-IDS2017 Dataset:

• 2,830,743 records with 79 features
• 308,381 duplicate rows (removed)
• 0.06% missing/infinite values (imputed via median)
• 47.5% memory reduction via numerical downcasting
• Severe class imbalance mitigation: 20% stratified sample

https://huggingface.co/papers/2604.14663#b-feature-engineering-and-selectionB. Feature Engineering and Selection

https://huggingface.co/papers/2604.14663#temporal-featuresTemporal Features

Δt_mean = (1/(n-1)) × Σ(t_i − t_{i-1})
Δt_std = √[(1/(n-1)) × Σ(Δt_i − Δt_mean)²]

https://huggingface.co/papers/2604.14663#entropy-based-featuresEntropy-Based Features

H(S) = −Σ p(s) log₂ p(s)

Captures distributional randomness in packet sizes.

https://huggingface.co/papers/2604.14663#feature-selectionFeature Selection

• Recursive Feature Elimination (RFE) using Random Forest permutation importance
• Ranking: I_j = (1/T) × Σ I(f_t(D) ≠ f^{-j}_t(D))

https://huggingface.co/papers/2604.14663#c-dimensionality-reduction-via-incremental-pcaC. Dimensionality Reduction via Incremental PCA

Cov(Z) = (1/(n-1)) × Z^T Z = V Λ V^T
Z_PCA = Z V_k

Result:Reduced from 78 to 35 principal components, retaining99.3% variancewhile reducing feature dimensionality by 55%.

https://huggingface.co/papers/2604.14663#d-class-balancing-strategiesD. Class Balancing Strategies

https://huggingface.co/papers/2604.14663#binary-classificationBinary Classification

• Random under-sampling: D_bal = D_min ∪ Sample(D_max, |D_min|)
• Result: 15,000 balanced instances (7,500 benign, 7,500 attack)

https://huggingface.co/papers/2604.14663#multi-class-classificationMulti-Class Classification

• SMOTE: x_new = x_i + λ(x_{ij} − x_i), where λ ~ U(0, 1)
• Adaptive SMOTE: λ ~ Beta(α, β), where α = 1 + ρ_i, β = 1 + (1 − ρ_i)

https://huggingface.co/papers/2604.14663#e-machine-learning-modelsE. Machine Learning Models

ModelConfigurationLogistic Regression (Elastic Net)α = 0.01, ρ = 0.5SVM (RBF Kernel)γ = 0.001, C = 1.0Random ForestT = 100 trees, max depth 20Gradient Boostingν = 0.1**Neural Network (MLP)**35 → 128 → 64 → K, dropout=0.5, Adam

https://huggingface.co/papers/2604.14663#f-evaluation-metricsF. Evaluation Metrics

• Accuracy, Precision, Recall, F1-Score
• Matthews Correlation Coefficient (MCC)
• Cohen’s Kappa (κ)
• Area Under ROC Curve (AUC-ROC)

https://huggingface.co/papers/2604.14663#5-experimental-setup5. Experimental Setup

https://huggingface.co/papers/2604.14663#a-dataset-construction-and-sampling-validationA. Dataset Construction and Sampling Validation

CIC-IDS2017 Sampling:

• Original: N = 2,830,540 flows
• Stratified 20% subset: n = 504,472
• Kolmogorov–Smirnov tests: p > 0.05 (no significant deviations)
• 92% of features: <5% mean deviation
• After PCA: k = 35 components (99.3% variance retained)

Train-Test Split:

• 80:20 stratified split (seed 42)
• Binary: 15,000 samples (7,500 benign, 7,500 attack)
• Multi-class: 35,000 samples via SMOTE (5,000 per class)

https://huggingface.co/papers/2604.14663#b-hyperparameter-optimizationB. Hyperparameter Optimization

Configurations:

• **Config 1 (Efficiency):**Computational efficiency prioritized
• **Config 2 (Expressiveness):**Accuracy maximized via 3-fold grid search

Key Hyperparameters:

• Logistic Regression: C ∈ {0.1, 100}
• SVM: RBF kernel with γ = 0.1
• Random Forest: n ∈ {100, 200}, depth=20
• Decision Tree: depth ∈ {6, 10, 15}
• KNN: k ∈ {3, 5, 7}

https://huggingface.co/papers/2604.14663#c-evaluation-protocolC. Evaluation Protocol

https://huggingface.co/papers/2604.14663#stage-1-cross-validationStage 1: Cross-Validation

• 5-fold stratified cross-validation on training partition (n = 12,000)
• Stratification preserves 50:50 benign-to-attack ratio
• Fold-to-fold variability: σ_CV = √[(1/(K-1)) × Σ(Acc_i − Acc̄)²]

https://huggingface.co/papers/2604.14663#stage-2-hold-out-testingStage 2: Hold-Out Testing

• Best configuration retrained on full training set
• Evaluated on held-out test set (n = 3,000, 20%)
• Metrics: Accuracy, Precision, Recall, F1, ROC-AUC, Confusion matrices

https://huggingface.co/papers/2604.14663#statistical-reliabilityStatistical Reliability

• Three independent random seeds: 42, 123, 456
• 95% confidence intervals: CI_95% = x̄ ± 1.96 × (σ/√n)

https://huggingface.co/papers/2604.14663#6-experimental-results6. Experimental Results

https://huggingface.co/papers/2604.14663#a-binary-classification-performanceA. Binary Classification Performance

https://huggingface.co/papers/2604.14663#linear-modelsLinear Models

• Logistic Regression: 92.21% accuracy (σ = 5.81 × 10⁻³)
• Config 2 improvement: +0.30% to 92.51%

https://huggingface.co/papers/2604.14663#kernel-based-methodsKernel-Based Methods

• SVM (Linear): 83.00% (underfits)
• SVM (RBF): 96.14% (+13.14%, σ = 3.89 × 10⁻³)

https://huggingface.co/papers/2604.14663#tree-based-ensemblesTree-Based Ensembles

• Random Forest Config 1: 95.98%
• Random Forest Config 2: 98.09% (+2.11%, σ = 1.72 × 10⁻³) ✓BEST

https://huggingface.co/papers/2604.14663#instance-based-learningInstance-Based Learning

• KNN (k=5): 97.40% (σ = 0.89 × 10⁻³)
• KNN (k=3): 97.93% (+0.53%, σ = 1.27 × 10⁻³)

https://huggingface.co/papers/2604.14663#b-multi-class-classification-performanceB. Multi-Class Classification Performance

ModelCV Acc.Test Acc.PrecisionRecallF1Random Forest (T=10, d=6)96.0±0.00997.196.997.096.9Random Forest (T=15, d=8, m=20)98.0±0.00798.097.998.0****97.9Decision Tree (d=10)96.0±0.01290.390.190.290.1KNN (k=7, distance-wt)94.0±0.01495.295.095.395.1

https://huggingface.co/papers/2604.14663#c-per-class-breakdown-random-forest-config-2C. Per-Class Breakdown (Random Forest Config 2)

Attack ClassPrecisionRecallF1-ScoreBENIGN99.2%98.5%98.9%DoS98.8%99.0%**98.9%**DDoS98.6%98.9%**98.7%**Port Scan95.7%97.6%**96.6%**Brute Force95.1%97.5%**96.3%**Web Attack91.9%96.0%**93.9%**Bot90.2%95.3%92.7%

https://huggingface.co/papers/2604.14663#7-federated-learning-convergence-analysis7. Federated Learning Convergence Analysis

https://huggingface.co/papers/2604.14663#a-convergence-and-compression-trade-offA. Convergence and Compression Trade-off

EdgeDetect achievesconvergence parity with full-precision FedAvgat 32× compression:

• Across 2.8M CIC-IDS2017 samples
• No measurable accuracy degradation (Δ < 0.2 pp)
• Cosine similarity: 0.87 ± 0.04

https://huggingface.co/papers/2604.14663#b-privacy-enhancement-through-smartificationB. Privacy Enhancement Through Smartification

MethodTechniquePSNR (dB)Label RecoveryFedAvg (Undefended)None31.7High-fidelitysignSGDZero-threshold16.8Partial recoveryEdgeDetectMedian-threshold15.1****14.3% (random)

https://huggingface.co/papers/2604.14663#c-theoretical-convergence-analysisC. Theoretical Convergence Analysis

Lemma 1 (Descent under Median-Threshold Smartification): Let L(W) be L-smooth and bounded below. Let g̃_t denote the smartified gradient with cosine similarity cos(θ_t) = ⟨g_t, g̃_t⟩ / (∥g_t∥ ∥g̃_t∥) ≥ γ > 0.

For sufficiently small step size η:

E[L(W_{t+1})] ≤ L(W_t) − ηγ∥g_t∥² + (Lη²/2)∥g̃_t∥²

Theorem 1 (Convergence under Bounded Variance): Assume bounded stochastic gradient variance σ² and cosine similarity γ > 0. Then after T rounds:

min_{t≤T} E[∥∇L(W_t)∥²] = O(1 / (γ√T))

https://huggingface.co/papers/2604.14663#8-federated-learning-scalability8. Federated Learning Scalability

https://huggingface.co/papers/2604.14663#a-convergence-under-different-heterogeneity-levelsA. Convergence Under Different Heterogeneity Levels

DistributionK=50 ClientsR₉₅R₉₈AccuracyBandwidthIIDFedAvg14228798.2%129.15 GBEdgeDetect14528998.0%4.05 GB**Non-IID (α=1.0)FedAvg20142396.4%190.35 GBEdgeDetect19239896.8%5.57 GBNon-IID (α=0.1)FedAvg31268793.8%309.15 GBEdgeDetect28761294.2%8.57 GBEdgeDetect+FedProx26456395.1%**7.88 GB

https://huggingface.co/papers/2604.14663#b-scalability-with-number-of-clientsB. Scalability with Number of Clients

K ClientsIID DistributionR₉₈AccuracyTotal Bandwidth10IID20198.1%2.81 GB25IID25498.0%3.56 GB100IID35697.9%4.98 GB500IID46797.7%6.54 GB **Sublinear scaling:**Increasing clients from K=10 to K=500 raises R₉₈ from 201 to 467 (sublinear in K).

https://huggingface.co/papers/2604.14663#9-ablation-study9. Ablation Study

https://huggingface.co/papers/2604.14663#component-wise-impact-analysisComponent-wise Impact Analysis

ConfigurationAccuracyCommunicationPSNR (dB)Invertible?Full EdgeDetect98.0%14.0 MB15.1No– Smartification98.2%450.0 MB ↑32×15.1Protected– Encryption (HE)98.0%14.0 MB31.7 ↑Yes– DP Noise98.1%14.0 MB14.2Protected– PCA (78 features)97.9%58.2 MB ↑4×15.3Protected– SMOTE94.2% ↓14.0 MB15.1ProtectedFedAvg (No Protection)98.2%450.0 MB31.7YessignSGD97.8%14.1 MB16.8Partial Key Findings:

• Smartification: Essential for communication efficiency (32×), negligible accuracy loss
• Encryption: Critical for privacy (PSNR 31.7 → 15.1 dB)
• SMOTE: Essential for accuracy (+3.8 pp gain)
• PCA: Reduces dimensionality (4.16×) with negligible impact

https://huggingface.co/papers/2604.14663#10-comparison-with-state-of-the-art10. Comparison with State-of-the-Art

StudyYearModelAccuracyF1DatasetClassesPrivacyComm. (MB)Centralized ApproachesAlam et al.2023CNN97.2%96.8CIC-IDS2017Binary✗N/AGhani et al.2023XGBoost96.1%95.4CIC-IDS20177-class✗N/ASavic et al.2021LSTM-AE95.5%94.2NSL-KDDBinary✗N/AFederated Learning ApproachesLiu et al.2023Fed-DNN96.3%95.1UNSW-NB155-classDP380Wang et al.2022Fed-CNN94.7%93.8CIC-IDS2017Binary✗520Zhang et al.2022FedAvg-LSTM93.5%92.4KDD-CUP994-classDP410Chen et al.2021Fed-XGB95.8%94.9IoT-23BinarySecAgg290This WorkEdgeDetect2026Fed-RF98.0%97.9%CIC-IDS20177-classHE****14(Binary)96.0%96.0%BinaryHE14 Key Advantages:

• Highest accuracyon CIC-IDS2017 (98.0% vs 96.3%)
• 96.9% communication reductionvs federated baselines (14 MB vs 290-520 MB)
• Strongest cryptographic privacy(Paillier HE vs DP/SecAgg)
• Practical edge deployment(4.2 MB, 0.8 ms on Raspberry Pi 4)

https://huggingface.co/papers/2604.14663#11-edge-deployment-evaluation11. Edge Deployment Evaluation

https://huggingface.co/papers/2604.14663#a-raspberry-pi-4-deploymentA. Raspberry Pi 4 Deployment

MetricRandom ForestKNNSVMLogistic Reg.Memory234 MB412 MB178 MB45 MBTraining Time12.3 s0.3 s*18.7 s2.4 sInference Latency****0.87 ms3.21 ms1.45 ms0.12 msEnergy per Inference12 mJ———Accuracy98.0%95.2%96.0%93.0% *KNN training is instantaneous (lazy learning) but requires 412 MB for storage.

https://huggingface.co/papers/2604.14663#b-resource-constrained-feasibilityB. Resource-Constrained Feasibility

• Memory footprint: 4.2 MB per client for gradient storage
• Encryption overhead: 156.4 ms per round (per-round encryption complexity O(d log n))
• Total bandwidth per round: 14 MB (vs 450 MB for full-precision)
• Accuracy loss on edge: <0.5% when deployed on Raspberry Pi 4

https://huggingface.co/papers/2604.14663#12-robustness-analysis12. Robustness Analysis

https://huggingface.co/papers/2604.14663#a-poisoning-attack-resilienceA. Poisoning Attack Resilience

**Setting:**5% to 20% of clients send poisoned updates

Poisoning RateAccuracyMacro F1p-value0% (Clean)98.0%0.979—5%96.4%0.961<0.00110%92.1%0.918<0.00115%89.3%0.887<0.00120%87.0%0.850<0.001 **Conclusion:**Maintains >85% accuracy even with 20% malicious clients (p < 0.001).

https://huggingface.co/papers/2604.14663#b-differential-privacy-utility-trade-offB. Differential Privacy-Utility Trade-off

εδAccuracyF1Privacy Loss10.010⁻⁵98.2%0.980Weak1.010⁻⁵98.1%0.979Moderate0.110⁻⁵96.8%0.965Strong

https://huggingface.co/papers/2604.14663#13-discussion13. Discussion

https://huggingface.co/papers/2604.14663#key-insights-for-federated-ids-in-6g-iotKey Insights for Federated IDS in 6G-IoT

1. PCA reveals strong redundancy: 35 components retain 99.3% variance with negligible performance loss, enabling efficient computation and communication.
2. Random Forest optimal: Best stability–accuracy trade-off (98.0% accuracy, 97.9% macro F1, σ = 0.0017).
3. Imbalance handling essential: SMOTE–undersampling improves minority recall from 0.39 to 0.98.
4. Gradient smartification superior to signSGD: - Preserves gradient alignment (0.87±0.04 cosine similarity) - Achieves 96.9% communication reduction - Improves privacy by lowering gradient entropy
5. Paillier encryption effective: Complete inversion resistance while retaining 98.7% of centralized accuracy.

https://huggingface.co/papers/2604.14663#challenges-and-future-workChallenges and Future Work

• Non-convex convergence: Theoretical analysis for deep learning architectures
• Concept drift: Adaptation to evolving attack patterns
• White-box robustness: Defense against adversarial gradient attacks
• Cumulative privacy loss: Formal composition under differential privacy

https://huggingface.co/papers/2604.14663#14-conclusion14. Conclusion

EdgeDetect introduces aprivacy-preserving federated intrusion detection frameworkfor resource-constrained 6G-IoT environments. The framework employs:

1. Gradient smartification: Median-based binarization achieving 32× communication reduction
2. Paillier homomorphic encryption: Only aggregated updates visible to server
3. Adaptive class balancing: SMOTE for minority-class robustness
4. Secure federated aggregation: Protection against inference and poisoning attacks

https://huggingface.co/papers/2604.14663#performance-summaryPerformance Summary

MetricValueMulti-class Accuracy98.0%Macro F1-Score97.9%Communication Reduction96.9% (450 MB → 14 MB)Edge Memory4.2 MBEdge Latency0.8 msPoisoning Resilience (20% attackers)87% accuracyGradient Inversion PSNR15.1 dB (vs 31.7 dB undefended) EdgeDetect demonstrates thatsecure federated IDS can meet strict privacy, efficiency, and reliability requirementsof next-generation 6G-IoT edge networks.

https://huggingface.co/papers/2604.14663#acknowledgmentsAcknowledgments

We thank the Canadian Institute for Cybersecurity for providing the CIC-IDS2017 dataset and the anonymous reviewers for their valuable feedback.

https://huggingface.co/papers/2604.14663#referencesReferences

[1] P. Kairouz, et al., “Advances and open problems in federated learning,” Foundations and Trends in Machine Learning, 2021.

[2] Y. Liu, J. Zhang, and H. V. Poor, “Federated deep learning for intrusion detection with differential privacy,” IEEE Transactions on Information Forensics and Security, 2023.

[3-62] [See original paper for complete reference list]

**Document Generated:**2026 Source:https://arxiv.org/abs/2604.14663v1