Outlier Detection in Incentivized Fault Tolerant Blockchain based Federated Machine Learning

Federated Machine Learning offers an exciting pathway for collaborative model training, enabling numerous users to contribute without disclosing their raw data. Yet, maintaining the security and privacy of both data and model within distributed settings continues to pose significant challenges. Identifying outliers plays a vital role in pinpointing abnormal behaviors that have the potential to weaken prediction accuracy or compromise the integrity of the model. The paper introduces a system that harnesses autoencoders in conjunction with anomaly scoring techniques and thresholding mechanisms to preemptively detect anomalies within the dataset prior to model training. The objective of the system is to optimize preprocessing stages by proactively filtering out potential breaches before data is introduced into the distributed environment. In the context of FML, where the model is trained across a distributed network, vulnerabilities arise as model parameters are exposed to evasion attempts. These attempts aim to undermine model integrity by manipulating the aggregation process. A protocol termed incentivized Probabilistic Byzantine Fault Tolerance is developed to ensure the integrity of the model during its training process in a distributed environment. The proposed framework offers a holistic solution to enhance security and integrity in distributed machine learning environment without compromising the system performance.