FlowerFED-IDS: A privacy-preserving IDS with quantified privacy-utility trade-offs

The rise of Internet of Medical Things (IoMT) and edge computing in healthcare demands intrusion detection systems (IDS) that protect sensitive patient data without sacrificing detection quality. Federated learning (FL) enables collaborative training of such detectors without sharing raw data; however, existing FL-based IDS remain vulnerable to model inversion and poisoning attacks during client–server communication, and the compared state-of-the-art solutions implement no formal privacy mechanism. This paper presents FlowerFED-IDS, a federated IDS built on the Flower framework that unifies three mechanisms in a single stack: a hybrid 1D CNN–BiLSTM detection model, local differential privacy (LDP) that clips and perturbs model updates before transmission, and robust FedTrimmedAvg aggregation. On the CICIoMT2024 dataset, under IID partitioning as a controlled baseline, FlowerFED-IDS achieves accuracies of 96.38%, 93.03%, and 93.89% with 5, 10, and 15 clients, respectively, maintaining recall above 93% in all configurations. Under a poisoning scenario with 20% compromised clients, FedTrimmedAvg keeps accuracy within 0.5 percentage points of the honest baseline. A privacy–utility analysis over noise multipliers identifies σ = 0.10 as the operating point satisfying the 95% clinical recall floor; we report its composed budget transparently (ε = 1215 at δ = 10 −5 over 20 rounds), a quantified but loose bound providing practical protection against update-reconstruction and inference attacks rather than a strict differential-privacy guarantee. A preliminary non-IID evaluation with Dirichlet α = 0.5, further confirms that the framework remains operational under realistic data heterogeneity, with recall above the clinical threshold, at the cost of reduced early-training stability.