FedREST-HIDS Notebooks and results accompanying "FedREST-HIDS: Split-Federated Intrusion Detection for Resource-Heterogeneous Industrial IoT" (Vyas, Hwang & Lin), submitted to Computer Communications. FedREST-HIDS partitions intrusion detection across three tiers. Resource-rich (RR) devices train complete local models and take part in federated learning. Resource-constrained (RC) devices execute only the initial feature-extraction layer and transmit clipped, noised activations to an edge server that runs the remaining layers. Federated transfer learning aligns the edge models globally, and a trust-aware aggregation rule (FedTrust) bounds the influence of any single client. Privacy is enforced locally under Rényi differential privacy, with the per-round budget converted once to an (ε, δ) guarantee.
Code accompanying "FedREST-HIDS: Split-Federated Intrusion Detection for Resource-Heterogeneous Industrial IoT" (Vyas, Hwang & Lin), submitted to Computer Communications. FedREST-HIDS partitions intrusion detection across three tiers. Resource-rich devices train complete local models and take part in federated learning. Resource-constrained devices execute only the initial feature-extraction layer and send clipped, noised activations to an edge server that runs the remaining layers. Federated transfer learning then aligns the edge models globally, and a trust-aware aggregation rule bounds the influence of any single client. Privacy is enforced locally under Rényi differential privacy.
Code accompanying "FedREST-HIDS: Split-Federated Intrusion Detection for Resource-Heterogeneous Industrial IoT" (Vyas, Hwang & Lin), submitted to Computer Communications. FedREST-HIDS partitions intrusion detection across three tiers. Resource-rich devices train complete local models and take part in federated learning. Resource-constrained devices execute only the initial feature-extraction layer and send clipped, noised activations to an edge server that runs the remaining layers. Federated transfer learning then aligns the edge models globally, and a trust-aware aggregation rule bounds the influence of any single client. Privacy is enforced locally under Rényi differential privacy.
