The increasing reliance on datasets in machine learning raises concerns about user privacy and data security. Federated Learning (FL) addresses this by client-side model training through contribution of model updates rather than raw data, enabling data ownership. Vulnerability of centralization risks, inference attacks and free-rider attacks persists necessitating the need for advanced security solutions. This paper proposes FedShield, a novel approach that integrates FL with Blockchain, Homomorphic Encryption (HE) and Zero-Knowledge Proof (ZKP). FedShield utilizes Blockchain for a transparent ledger, HE for computations on encrypted data eliminating need for decryption, and ZKP for verifying model updates while maintaining confidentiality. Clients train local models on their datasets, and encrypt the updates which are verified before storage. Encrypted local model updates are homomorphically averaged to generate an aggregated global model update, which boosts local model generalization and recommendation. FedShield achieves over 90% accuracy in a video recommender system while ensuring privacy preservation.