Malicious domain detection (MDD) from DNS telemetry enables early threat hunting but is constrained by privacy and data-sharing barriers across organizations. We present a deployable federated learning (FL) pipeline that trains a compact deep neural network (DNN; 64-32-16 with ReLU and dropout 0.3) locally at each client and exchanges only masked model updates. Privacy is enforced via secure aggregation (the server observes only an aggregate of masked updates) and optional server-side differential privacy (DP) via clipping and Gaussian noise. Our feature schema combines DNS-specific lexical cues (character $n$-grams, entropy, TLD indicators) with lightweight behavioral signals (TTL dispersion, query cadence) without exporting raw logs or identifiers. We benchmark FedAvg, FedProx, and FedNova under controlled non-IID client partitions and report ROC-AUC, precision-recall area under the curve (PR-AUC), F1, convergence speed, and communication cost. Federated models approach centralized training while outperforming local-only baselines; FedProx reaches the target Accuracy $\ge 0.995$ in fewer rounds than FedAvg under medium heterogeneity. We report 95% bootstrap confidence intervals and paired significance tests (DeLong for ROC-AUC; McNemar for Accuracy). Overall, privacy-preserving FL for DNS-based MDD is practical, providing near-centralized utility while keeping DNS data local.