With the increasing autonomous capabilities of distributed cyber-physical systems, the complexity of their models also increases significantly, thus continually posing challenges to existing formal methods for safety verification. In contrast to model checking, monitoring emerges as an effective lightweight, yet practical verification technique capable of delivering results of practical importance with better scalability. Monitoring involves analyzing logs from an actual system to determine whether a specification (such as a safety property) is violated. Monitoring techniques, such as those using reachability methods, may fail to produce results when dealing with complex models like Deep Neural Networks (DNNs). We propose here a novel statistical approach for monitoring that is able to generate results with high probabilistic guarantees. We evaluate our monitoring technique on three case studies.