Privacy Enhancement using Federated Learning & Blockchain
Machine learning pipelines executed within cloud infrastructure, are computationally intensive & use specialized hardware utilizing parallel processing. These systems encounter challenges in domains of cost reduction, scalability, reliability & efficiency, real-time analysis & prediction, security & privacy.
Federated Learning is a decentralized, collaborative approach where machine learning models are trained across multiple devices or servers. It helps alleviate systemic privacy risks & costs resulting from traditional, centralized machine learning. It employs secure multi-party computation, which is a tool for input privacy, wherein multiple parties have access to the output of a function while keeping the inputs secret from each other. The model lives on the device & learns from the data on the device without uploading any information to a central server. Model updates are integrated with model on the cloud. What this means is that instead of distributing data across devices to train the model, the model is distributed across devices & updates from all devices are integrated to form an updated model on a central server. A central server and client devices are the two main participants in a federated learning system. At every stage of the federated learning process there are mandatory and optional components. The process of federated learning begins with job creation, followed by, data collection & pre-processing, model training, deployment & monitoring.
Classical federated learning has some drawbacks which include security issues such as data attacks, single-point failures, communication delays between remote & central servers, non-transparent model aggregation & lack of incentive mechanisms. They operate under the conditions that all devices & server establish a strong connection to train models collaboratively & devices have to trust the central server. These challenges can be mitigated by combining federated learning with blockchain.
Blockchains are decentralized ledgers that record transaction information in chronological order & are immutable. With blockchain, federated learning gets added on with a trusted consensus mechanism which allow trusted exchange of updates from models among connected devices. This eliminates the requirement of a central server & risks of single-point-of failures. There is trust between devices & servers, improved scalability of intelligent edge networks. Its limitations are possibilities of latency, energy cost required for blockchain mining & possible privacy risks from sharing transaction updates. A possibility of conflict between training data on organization & blockchain exists. The connected devices train learning models & run mining for extra profit. Replacing the central server with block chain eliminates the risk of single-point failure in the system.
For example, unmanned aerial vehicle networks on chain have improved security against adversarial attacks as the shared ledgers cannot be altered. All peer UAVs have the transparency to monitor channel behavior and are traceable among network entities
The combination of federated learning and blockchain provides a trusted ecosystem that would enable advancements in development of smart cities.
