IoT devices stop trusting a one-time credential: behavioral fingerprints plus federated learning cut membership inference success from 48.2% to 12.4%

IoT authentication becomes continuous verification of a device's own behavior rather than a single credential check followed by implicit trust. Static credential methods cannot detect abnormal behavior after a device passes initial authentication, centralized trust has a single point of failure, and identifiable attributes stay exposed during communication. On 11 Raspberry Pi devices in the PP-IDE behavioral device fingerprinting dataset, a Random Forest classifier identified devices with 99.21% accuracy, a 1.2% false acceptance rate and a 0.9% false rejection rate.

Privacy comes from training locally and sharing only noised model updates, so raw behavioral data never leaves the device. Federated Averaging aggregates local updates, and differential privacy adds noise with a noise multiplier of 1.0, gradient clipping at 1.0, ε = 5.0 and δ = 1×10⁻⁵. Under IID and non-IID settings, membership inference attack success fell from 48.2% to 12.4% and from 45.1% to 11.8% respectively, with 95.12% accuracy in the non-IID differential privacy setting.

Authentication records go to a lightweight permissioned chain, and multiple agents coordinate authentication decisions across edge, fog and cloud layers. Consensus uses a reputation-based lightweight proof-of-authority mechanism instead of traditional consensus with high energy and latency cost; behavioral assessment, federated learning, blockchain verification and orchestration agents fuse signals through attention. The chain reached 1,200 transactions per second, 1.2 s authentication latency and $0.006 per authentication; agentic orchestration averaged a 97.8% detection rate and 98.4 ms response time.

Evaluation was mainly done on 11 Raspberry Pi devices in the PP-IDE dataset, with attack resistance drawing on network traffic from CIC-IoT-2023, and experiments ran on a simulated testbed of IoT devices, edge nodes, fog servers and cloud infrastructure. The authors propose further testing on more IoT devices, TinyML optimization for ultra-low-power devices, blockchain deployment at thousands of devices, and adaptability to hybrid attacks.

The authors note that a simulated environment may not accurately reflect the complexity of real IoT deployments, that the explanation fidelity assessment was small-scale and qualitative, and that the blockchain module has not been tested at thousands of devices. The federated learning module still needs compute on resource-constrained devices, which the authors propose addressing with TinyML.