Anomaly detection for the Internet of Things (IoT) is a very important topic in the context of cyber-security. Indeed, as the pervasiveness of this technology is increasing, so is the number of threats and attacks targeting smart objects and their interactions. Behavioral fingerprinting has gained attention from researchers in this domain as it represents a novel strategy to model object interactions and assess their correctness and honesty. Still, there exist challenges in terms of the performance of such AI-based solutions. The main reasons can be alleged to scalability, privacy, and limitations on adopted Machine Learning algorithms. Indeed, in classical distributed fingerprinting approaches, an object models the behavior of a target contact by exploiting only the information coming from the direct interaction with it, which represents a very limited view of the target because it does not consider services and messages exchanged with other neighbors. On the other hand, building a global model of a target node behavior leveraging the information coming from the interactions with its neighbors, may lead to critical privacy concerns. To face this issue, the strategy proposed in this paper exploits Federated Learning to compute a global behavioral fingerprinting model for a target object, by analyzing its interactions with different peers in the network. Our solution allows the training of such models in a distributed way by relying also on a secure delegation strategy to involve less capable nodes in IoT. Moreover, through homomorphic encryption and Blockchain technology, our approach guarantees the privacy of both the target object and the different workers, as well as the robustness of the strategy in the presence of attacks. All these features lead to a secure fully privacy-preserving solution whose robustness, correctness, and performance are evaluated in this paper using a detailed security analysis and an extensive experimental campaign. Finally, the performance of our model is very satisfactory, as it consistently discriminates between normal and anomalous behaviors across all evaluated test sets, achieving an average accuracy value of 0.85.