Quantized Hybrid Privacy Preserving Approach for Federated Learning with Flower Framework

The increasing adoption of Decentralized AI systems, especially Federated Learning, has raised significant concerns regarding the privacy and security of sensitive data. These issues can lead to severe financial losses with breaches costing organizations millions annually. To address these issues, this study proposed the Quantized Hybrid Privacy Preserving (QHPP) approach for enhancing data security by maintaining the balance between accuracy, privacy level and efficiency. This novel approach ensures optimal trade-off between accuracy, privacy and efficiency by the unique integration of hybrid techniques with quantization. The study utilized MNIST and Employee Attrition datasets, along with simulation of attacks like label flipping, data poisoning and backdoor attack. Various privacy preserving techniques like Secure Aggregation, Homomorphic Encryption and Differential Privacy along with hybrid approach, combining all the techniques are evaluated using metrics such as accuracy, privacy level, computational and communicational overhead. The results indicate that the hybrid model achieved the highest accuracy of 94.73% across all attacks but suffered from high computational and communicational overhead, highlighting its inefficiency. To address this issue, optimization techniques were evaluated with quantization outperforming other techniques. The proposed QHPP achieved 94.57% accuracy and 0.97 privacy level with significantly lower overhead 1622.11 sec and 6.35 MB on MNIST and 1982 sec and 7.13 MB on Employee Attrition, highlighting its efficiency. Enterprises and Organizations will benefit from the QHPP approach for its efficiency, reduced overhead and enhanced privacy level. The future work will include scalability analysis of QHPP in large-scale FL systems with different datasets and more clients.