Personalized food recommendation systems, due to their reliance on centralized data storage, often suffer from privacy and security vulnerabilities. This research proposes a privacy-preserving framework integrating functional encryption and adversarial learning to safeguard sensitive user information while maintaining recommendation accuracy. Additionally, Fernet encryption is used for data encryption. A data poisoning attack is simulated to assess system resilience by injecting adversarial noise into encrypted datasets. Data imbalance is addressed using the Synthetic Minority Oversampling Technique (SMOTE) to enhance model fairness. Experimental results demonstrate that the proposed system effectively mitigates adversarial threats, preserves user privacy, and achieves robust performance across various metrics. This work advances secure and privacy-aware recommendation systems, offering a practical approach to defending against data manipulation attacks.