Visual simultaneous localization and mapping (SLAM) is fundamental to autonomous mobile robots, yet transmitting raw visual features between distributed agents or to cloud backends exposes sensitive scene data to interception and reconstruction attacks. We introduce SecureSLAM, a privacypreserving visual SLAM framework that performs encrypted feature aggregation over learned keypoint descriptors without sacrificing localization accuracy. Our approach integrates three novel components: (i) a lightweight homomorphic encryption layer that operates directly on SuperPoint-style deep feature descriptors, (ii) a secure loop-closure verification protocol based on privacy-preserving approximate nearest-neighbor search, and (iii) a federated map-merging strategy that enables multi-robot collaboration without sharing raw observations. Evaluated on the KITTI Odometry, EuRoC MAV, and TUM RGB-D benchmarks, SecureSLAM achieves translational errors within 3–7% of unencrypted baselines while providing formal (ϵ, δ)-differential privacy guarantees on all transmitted features. On KITTI, our system attains a mean translational RMSE of 0.42 m compared to 0.39 m for the plaintext oracle, while reducing feature inversion attack success from 87.3% to 4.1%. These results demonstrate that strong privacy protection and accurate visual SLAM are not mutually exclusive, opening a path toward secure deployment of assistive and industrial mobile robots in privacy-sensitive environments.