Recent studies have shown that gradient updates in federated learning (FL) can unintentionally expose sensitive information about a client’s local data. When the server is malicious and manipulates the global model, the risk escalates because the server can provoke information‑rich updates from clients. This paper adopts a defender’s perspective to deliver the first systematic analysis of malicious gradient‑leakage attacks and the model‑manipulation techniques that enable them. Our investigation uncovers a fundamental trade‑off: in realistic FL settings that include common normalization methods and federated averaging, an attack cannot be both highly effective at reconstructing private data and sufficiently stealthy to avoid detection. Consequently, while theoretically concerning, malicious gradient leakage is practically limited and often detectable through basic monitoring. To address this, we propose a simple, lightweight client‑side detection mechanism that inspects model updates before local training begins and flags suspicious ones. The approach is easy to implement, incurs minimal overhead, and offers a deployable safeguard for privacy‑sensitive FL systems, even though detection may not be strictly necessary in some realistic scenarios.\n\nReview: By highlighting the inherent tension between attack efficacy and stealth, the work provides a realistic assessment of gradient‑leakage threats and supplies a practical defense, advancing the security of federated learning deployments.