Financial institutions face major challenges with regards to data privacy, fraud detection, and credit risk assessment as a result of the weaknesses in centralized AI systems that deal with sensitive financial data. These are susceptible to cyberattacks, breaches of privacy, and inability to comply with regulation. In an effort to solve these challenges, we propose FinSec-FL, a robust Federated Learning (FL) framework to enhance secure, scalable, and privacy-oriented AI in financial services. It leverages homomorphic encryption, secure multiparty computation, and differential privacy methods that allow cooperative AI model training while preserving raw data in confidence. Moreover, it includes blockchain-smart contracts for maintaining some model verification and auditability, which is also transparent and tamper-resistant for confidential information. This method adds higher resilience of the system against malicious activity and trust for the decentralized financial networks. The framework contains a method known as Adaptive Federated Learning, which facilitates rapid fraud detection, and also federated reinforcement learning to handle credit risk in dynamic environments. This research proves that FinSec-Fl is more accurate than all other centralized models for fraud detection, which protects privacy and computational efficiency, which is crucial. Future work would involve quantum-resistant cryptography and federated graph neural networks to further detect neural networks to detect fraud patterns, which is the future objective of the project.