Cyber-Physical Systems (CPS) like smart grids, industrial automation units, and modern manufacturing plants are associated with the need to make decisions in real-time and ensure high security. Machine Learning (ML) has strong detection and prediction functions, yet the conventional inference with the help of a CPU or cloud is too slow and can be compromised. At the same time, the integrity of CPS is severely endangered by compromised ML models or bitstreams of the FPGA. In this paper, the author suggests a FPGA-Accelerated Machine Learning Framework with a Blockchain-Based Integrity Validation as a means of securing CPS environments. The method employs FPGAs to detect anomalies and predictive models at the edge and attains ultra-low latency. A blockchain layer provides tamper-proof model provenance, validation of bitstream and safe event logging. The integrated design is both very reliable, very fast, and decentralizes trust, and is resistant to model pollution and device impersonation. The experimental evidence of using representative ML models illustrates a substantial decrease in the time of inference, latency, and security robustness.