Federated learning (FL) enables collaborative Intrusion Detection Systems (IDS) across distributed Internet of Things (IoT) networks without sharing raw data. However, its openness exposes it to model poisoning and backdoor attacks, where malicious clients manipulate updates to corrupt the global model. Detecting such threats remains difficult under non-independent and identically distributed (non-IID) data, where benign updates diverge naturally and resemble adversarial behavior. Existing defenses based on robust aggregation, clustering, or cryptographic auditing either assume IID conditions, lack a global semantic reference, or incur high computational costs, leading to poor robustness and high false positives. To overcome these challenges, we propose probe-for-poisoning (P4P), a lightweight system-level integration and refinement of similarity-based defenses for securing FL aggregation. P4P constructs a probe vector from the global model’s historical trajectory and actively queries clients. To establish trust, the server itself computes a single scalar probe response representing the cosine similarity between each client's received local gradient and the probe vector, ensuring all computationally intensive detection is performed centrally without relying on client-reported metrics. Combined with L2-norm filtering, ensemble anomaly detection, and temporal suspicion scoring, P4P forms a multi-metric defense pipeline capable of identifying both abrupt and stealthy poisoning behaviors. Experiments on IoTDIAD 2024 and ACI-IoT 2023 datasets show that P4P improves accuracy by up to 18\% and reduces false positives by 25\% compared to state-of-the-art defenses such as FedMP and Recess, with negligible overhead. These results highlight P4P as a robust and practical anti-poisoning solution for federated intrusion detection under realistic, resource-constrained IoT network scenarios.