Source-linked AI summary
Byzantine-Robust Federated Fire Detection with a Rotating Coordinator
Georgia Argyrou, Aymen Bahrouny, Hedi Fendriy, Alexander Jung
TL;DR
Indoor fire-detection FL must preserve camera privacy while addressing bandwidth limits, Byzantine clients, and trust in a fixed coordinator. The paper combines a curated dataset, compressed edge updates, history-aware aggregation, and coordinator rotation; its experiments show stealthy attacks can be evicted and the approach remains feasible in a distributed deployment.
Problem
Federated indoor fire detection must handle privacy-sensitive distributed cameras, limited uplink bandwidth, Byzantine clients, and the single point of failure created by a permanently fixed coordinator.
Method
The paper builds an 18,790-image dataset, a frozen-MobileNetV2 detector with compressed head updates, and SafeguardSGD on a rotating coordinator that preserves client detection state.
Results
The method evicts stealthy attacks that per-round filters miss, compresses uplinks up to 10×, and matches its fixed-server counterpart in reported balanced accuracy and detection speed.
Takeaways & Limitations
The experiments support a semi-decentralized federated fire-detection design that combines communication reduction, Byzantine robustness, and distributed coordination.
Takeaways & Limitations
Evidence is empirical and limited to five clients with a single static, non-colluding adversary.
Abstract
from arXiv · showhide
We study the application of federated learning (FL) to indoor fire detection. Such fire-detection systems use edge cameras that record sensitive footage which cannot easily be collected at a central server. Existing federated solutions leave three practical obstacles unaddressed: limited uplink bandwidth, Byzantine (malicious or faulty) clients, and unconditional trust in a single, permanently fixed aggregation server. Our main contributions address all three. In particular, we provide (i) a curated indoor fire-detection dataset assembled from eight public sources; (ii) an edge-deployable detector whose model updates are compressed up to 10 time with only a small loss in balanced accuracy; and (iii) a semi-decentralized Byzantine-robust FL method that combines history-aware aggregation with a rotating coordinator, evicting stealthy attacks that per-round filters miss while removing the fixed-server single point of failure. On the held-out test set the rotating-coordinator method matches its fixed-server counterpart in accuracy and detection speed, and a physically distributed six-node cloud deployment confirms feasibility.
1. INTRODUCTION
Indoor fire cameras can detect visual cues before conventional sensors activate, but federated deployment must handle privacy, heterogeneous clients, bandwidth limits, Byzantine updates, and reliance on one fixed coordinator.
- Indoor fires spread rapidly, while conventional smoke and heat detectors can activate late and produce false alarms.
- Federated learning keeps raw camera data on-device while exchanging model updates, fitting privacy-sensitive fire surveillance across separately owned buildings.
- Federated fire detection must address non-IID client data, arbitrarily corrupted Byzantine updates, expensive full-update transmission, and a fixed coordinator’s single point of failure.
- The paper contributes an 18,790-image indoor dataset assembled from eight public sources for IID and non-IID federated evaluation.
- The edge detector compresses head updates up to 10× using INT8/INT4 quantization and error-feedback Top-K sparsification.
- SafeguardSGD combines history-aware aggregation with a rotating coordinator, preserving detection state while evicting stealthy attacks that per-round filters miss.
2. SYSTEM AND DATA
The system uses a curated indoor fire dataset and a lightweight frozen-backbone detector, with federated evaluation under IID and non-IID client partitions.
- The 18,790-image binary dataset combines eight public sources, is deduplicated and cleaned, and spans homes, offices, warehouses, and corridors.
- Training and validation data are distributed across five clients under IID and approximately 70% majority-class non-IID label-skew splits.
- A frozen ImageNet-pretrained MobileNetV2 feeds a trainable 1280 →256 →2 head, allowing cached features and lightweight federated optimization.
- Only the head’s N = 328,450 parameters are exchanged, reducing the federated communication payload relative to full-model training.
- FedAvg reaches test BA 0.9835±0.0014 on IID data and 0.9818±0.0026 on non-IID data, with fire recall ≈0.98.
3. COMMUNICATION-EFFICIENT AGGREGATION
Message-level quantization and sparsification reduce uplink payloads while retaining most balanced accuracy, with gradual degradation as compression increases.
- 0.9754 BA at 10× compression compares with 0.9835 uncompressed, an approximately 0.8-point gradual decline in the communication trade-off.
- INT8 at 4× reduces BA by only 0.0016, from 0.9835 to 0.9819.
- INT4 at 8× and error-feedback Top-K with k = 0.05 at 10× reduce BA to 0.9782 and 0.9754, respectively.
- The 328,450-parameter head shrinks from a 1283 KB float32 payload to 128 KB, while encoding takes ≤17 ms per round versus approximately 30 s of local training.
- Table 1 reports per-client, per-round payload, compression ratio over float32, and global test balanced accuracy.
4. HISTORY-AWARE BYZANTINE ROBUSTNESS
The Byzantine-robustness evaluation studies one static Byzantine client among five and compares history-aware permanent eviction with stateless per-round defenses.
- The evaluation uses five clients, at most one Byzantine client, a non-omniscient static adversary, and attacks ranging from overt divergence to stealthy persistent bias.
- Krum and centered clipping are per-round defenses that do not permanently evict clients and can miss geometrically central or below-threshold attacks.
- SafeguardSGD accumulates deviations over short and long windows, crossing τ = 0.75 for persistent Byzantine behavior while a norm floor ηfloor = 0.5 limits premature warm-up eviction.
- SafeguardSGD evicts stealth-flip attackers for α ≥0.3, with mean eviction rounds falling from 28 at α = 0.3 to 10 at α = 0.8.
5. ROTATING-COORDINATOR ARCHITECTURE
The rotating coordinator transfers aggregation state across edge nodes to remove dependence on a permanently fixed server while preserving Byzantine-defense behavior. In deployment, the system achieved identical reported balanced accuracy under fixed and rotating architectures.
- Byzantine robustness: SafeguardSGD permanently evicts stealth-flip attackers faster as perturbation size grows, while fixed and rotating variants overlap in eviction timing.The figure compares eviction rounds for both architectures and shows Krum continuing to select malicious updates.
- Coordinator rotation: Coordinator rotation serializes global weights, client accumulators, retained clients, participation history, and the next coordinator identifier at each handoff.Preserving these states allows SafeguardSGD’s detection state to survive role transfer.
- Evaluation: Table 3 reports test balanced accuracy and mean eviction round for fixed versus rotating SafeguardSGD across attacks.These metrics directly evaluate whether rotation preserves model quality and Byzantine eviction behavior.
- Coordinator rotation: IPM BA 0.9545 under both fixed and rotating architectures, while a fixed server never exposes the attacker to coordinator election.The reported deployment used six distributed EC2 instances, with one rotated coordinator and five clients.
6. CONCLUSION
The paper presents a federated indoor fire-detection system addressing communication cost, Byzantine clients, and dependence on a fixed coordinator. Across an 18,790-image dataset and distributed deployment, it reports preserved balanced accuracy, up to 10× uplink reduction, and empirical tolerance of a Byzantine client.
- Conclusion: The system matches the centralized baseline’s balanced accuracy, reduces uplink up to 10×, and tolerates a Byzantine client without a fixed trusted server.The evidence is empirical and uses five clients with one static, non-colluding adversary.
- Conclusion: History-aware SafeguardSGD permanently evicts stealthy low-magnitude attacks that per-round geometric filters miss, while rotation distributes aggregation across edge nodes.The rotating architecture preserves the Byzantine-defense behavior while removing the fixed-server single point of failure.