Quantization-Aware Federated Learning for Resource-Constrained IoT Intrusion Detection: Accuracy, Communication and Robustness Trade-Offs

Federated learning (FL) allows Internet of Things (IoT) gateways to train a shared intrusion detection system (IDS) without exporting raw network traffic, yet three constraints govern whether it is practical at the network edge: the memory and compute budget of the device, the capacity of the uplink, and exposure to poisoned updates from compromised clients. Existing studies largely address these constraints one at a time. This paper presents a systematic study of quantization-aware federated learning (QAFL) for IoT intrusion detection that jointly varies the on-device weight precision (32, 8 and 4 bits), the uplink update precision (32, 8, 4 and 2 bits, with and without error feedback), the server aggregation rule and the degree of statistical heterogeneity across clients. We also examine fixed-range update quantization, in which a server-declared quantization range bounds the influence of any single client as a side effect of compression. Across 876 training runs on three public benchmarks (CICIoT2023, Edge-IIoTset and TON_IoT) under Dirichlet non-IID partitioning, 8-bit weights with 4-bit updates retain 99.7% of full-precision macro-F1 (58.81 versus 58.98) while reducing per-round uplink traffic by 8×, model memory by 3.6× and estimated inference energy by 4.8×, and post-training quantization proves as effective as quantization-aware training for these compact models. Two-bit stochastic updates match full-precision accuracy with 15.9× less uplink traffic, but only without error feedback, which reduced 2-bit macro-F1 by 5.5 points; diagnostic runs trace this loss mainly to error-feedback memories that grow to 9 to 18 times the size of the update, with stale memories under partial participation adding to it. Fixed-range quantization reduced the macro-F1 loss caused by 20% sign-flipping clients from 59.1 to 3.8 points with plain FedAvg, at no cost in clean accuracy. A control with full-precision coordinate clipping at the same threshold gives the same protection, so the benefit comes from the server-set range rather than from the quantization grid. The defense bounds the magnitude but not the direction of malicious updates: a worst-case attacker that stays within the range attains this bound and defeats plain FedAvg from 20% to 25% attackers, and the distance-from-median filter of the closest prior framework is more robust whenever the server can inspect individual updates. Fixed-range quantization is thus a zero-cost, format-enforced bound that complements robust aggregation rather than replacing it. We consolidate these results into configuration guidelines for resource-constrained IoT deployments. All experiments run on a single commodity cloud GPU.

Authors

Institutions

Publication Details

Journal
Future Internet
Published
2026-11-18
DOI
https://doi.org/10.3390/fi18110638
Primary Topic
Privacy-Preserving Technologies in Data
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Quantization-Aware Federated Learning for Resource-Constrained IoT Intrusion Detection: Accuracy, Communication and Robustness Trade-Offs

Sumendra Yogarayan, M. Lakshmi, Mohan Muralidhar, Rubini Pandu et al.
3 links
Future Internet
Privacy-Preserving Technologies in Data
article

Quantization-Aware Federated Learning for Resource-Constrained IoT Intrusion Detection: Accuracy, Communication and Robustness Trade-Offs

Sumendra Yogarayan, M. Lakshmi, Mohan Muralidhar, Rubini Pandu, Dirasavancha Raghunath Kumar Babu, Zakea Ilagure
article en

Abstract

Federated learning (FL) allows Internet of Things (IoT) gateways to train a shared intrusion detection system (IDS) without exporting raw network traffic, yet three constraints govern whether it is practical at the network edge: the memory and compute budget of the device, the capacity of the uplink, and exposure to poisoned updates from compromised clients. Existing studies largely address these constraints one at a time. This paper presents a systematic study of quantization-aware federated learning (QAFL) for IoT intrusion detection that jointly varies the on-device weight precision (32, 8 and 4 bits), the uplink update precision (32, 8, 4 and 2 bits, with and without error feedback), the server aggregation rule and the degree of statistical heterogeneity across clients. We also examine fixed-range update quantization, in which a server-declared quantization range bounds the influence of any single client as a side effect of compression. Across 876 training runs on three public benchmarks (CICIoT2023, Edge-IIoTset and TON_IoT) under Dirichlet non-IID partitioning, 8-bit weights with 4-bit updates retain 99.7% of full-precision macro-F1 (58.81 versus 58.98) while reducing per-round uplink traffic by 8×, model memory by 3.6× and estimated inference energy by 4.8×, and post-training quantization proves as effective as quantization-aware training for these compact models. Two-bit stochastic updates match full-precision accuracy with 15.9× less uplink traffic, but only without error feedback, which reduced 2-bit macro-F1 by 5.5 points; diagnostic runs trace this loss mainly to error-feedback memories that grow to 9 to 18 times the size of the update, with stale memories under partial participation adding to it. Fixed-range quantization reduced the macro-F1 loss caused by 20% sign-flipping clients from 59.1 to 3.8 points with plain FedAvg, at no cost in clean accuracy. A control with full-precision coordinate clipping at the same threshold gives the same protection, so the benefit comes from the server-set range rather than from the quantization grid. The defense bounds the magnitude but not the direction of malicious updates: a worst-case attacker that stays within the range attains this bound and defeats plain FedAvg from 20% to 25% attackers, and the distance-from-median filter of the closest prior framework is more robust whenever the server can inspect individual updates. Fixed-range quantization is thus a zero-cost, format-enforced bound that complements robust aggregation rather than replacing it. We consolidate these results into configuration guidelines for resource-constrained IoT deployments. All experiments run on a single commodity cloud GPU.

Future InternetVol. 18(11)
Higher Colleges of Technology (AE), Nitte University (IN), Multimedia University (MY), B.M.S. College of Engineering (IN), CMR University (IN), Nitte Meenakshi Institute of Technology (IN)
Privacy-Preserving Technologies in Data
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon