A Blockchain-Based Model Migration Approach for Secure and Sustainable Federated Learning in IoT Systems

[Audio] Hi everyone today we aer going to talk about the paper ...

1 Introduction. 2 Related Work. 3 Methodology & Results.

1 Introduction.

Introduction. 1. Proposes a novel approach to improving the security and sustainability of federated learning in Internet of Things ( IoT ) systems . The proposed approach uses blockchain technology to store and manage model updates, which enhances security and transparency, and a consensus mechanism to ensure that all nodes in the federated learning system agree on the validity of the model updates ..

Introduction. 1. The approach uses blockchain technology to securely share models trained by fast devices and enhance security and transparency in the model migration process. The proposed approach aims to improve the efficiency and sustainability of federated learning on IoT devices, which could have important practical applications in various industries and domains. Federated learning is a machine learning technique that allows multiple devices to collaboratively train a model without having to share their data with a central server..

Related Work. A lgorithm proposed. 2. Reduce the training cost and achieve sustainable federated learning in edge computing and IoT environments. The authors propose an algorithm for learning model parameters from data distributed across multiple edge nodes while preserving privacy. The algorithm determines the best tradeoff between local update and global aggregation to minimize the loss function in resource-constrained settings..

Related Work. Federated Learning Method. Federated learning is a distributed machine learning method that addresses the problem of private data leakage. Traditional machine learning schemes gather data to a central server, which can result in privacy issues. The authors propose an ε-accuracy loss metric to measure the effectiveness of federated learning algorithms and compare their performance to traditional distributed learning methods..

3 Methodology & Results.

Blockchain. Blockchain. 3. Blockchain is a distributed ledger that aims to establish trust among nodes without needing a trusted central party.It uses a smart contract, which is an automatically executing, Turing complete, event-driven program code running on the blockchain . There are three types of blockchains : public, private, and consortium. Public blockchains are decentralized and open to all nodes, while private blockchains are mostly used inside organizations as a distributed database..

PROBLEM STATEMENT. @ 0.0) (0.132. 0) (0. B2. [33).

Threat Model. “. ”. IoT devices may deviate from the model migration approach by not training models or training them at a slower speed, or by committing fraud for rewards. The authors assume that IoT devices are rational and may fake feature extractors instead of real training models for quick feedback to the central server. However, the authors assume that IoT devices are on alert and can regard the blockchain as a trusted infrastructure , with attacks on the blockchain itself not being a consideration..

Challenges and Design Goals. “. Overcoming Hurdles and Achieving Objectives: Towards a Secure and Sustainable Federated Learning Model Migration Approach.

Node and Edge Encoding. The proposed blockchain -based secure federated learning model migration approach accelerates federated learning training speed and ensures security by dividing IoT devices into fast and slow, sharing feature extractors, and verifying local models' correctness..

Phases of the Proposed Approach. Initialization: The central server deploys the blockchain system and generates an original global model based on its local data. The server estimates the training time threshold T¯ of the fast devices based on device information collected in advance, and records these system parameters in the smart contract. Verify and publish feature extractor: When the current round of training starts for T¯ time, the server verifies the feedback received from all fast devices to prevent fraudulent behavior. The server selects the fastest device from the current fast devices and sends the feature extractor of its local model to slow devices with a similar data distribution through the on-chain channel..

04. E XPERIMENTAL E VALUATION AND T HEORETICAL A NALYSIS.

Analyzing and Testing the Proposed Blockchain -based Federated Learning Approach.

Analyzing and Testing the Proposed Blockchain -based Federated Learning Approach.

1 Simulation Settings.

Simulation Settings. Sma 42. Simulation Settings.

2 Functional Evaluation.

Functional Evaluation. Functional Evaluation. 2. 'raining (b).

3 Security Evaluation.

Security Evaluation. 2. Nomal Devices Malicious Devices Nodes (a).

Security Evaluation. Security Evaluation. Adjacency Matrix.

4 Performance Evaluation.

Performance Evaluation. "Cost Analysis of Blockchain -based Federated Learning Model Migration Approach for IoT Devices." T he authors evaluate the improvements in computation and storage costs of IoT devices brought by the proposed approach. They mainly evaluate the overhead of on-chain operations, which require a certain amount of gas in the Ethereum platform. Storing a 256-bit integer on the blockchain requires 20000 units of gas, and all on-chain operations consumed less than 60000 units of gas. The main overhead is storing data permanently on the blockchain , but the computation cost of calculating the reward for the device based on the incentive function is acceptable..

5 Conclusion.

Summary. “. Conclusion. The article proposed a blockchain -based federated learning approach for secure and sustainable IoT devices. The approach includes an incentive mechanism and uses the blockchain to protect security. The theoretical analysis and experiments show the approach's effectiveness, security, and sustainability. Future work includes designing a sampling strategy and optimizing blockchain usage to reduce computational costs..

Y. Wu, “Cloud- edge orchestration for the Internet of Things : Architecture and AI- powered data processing ,” A. Imteaj , U. Thakker , S. Wang, J. Li, and M. H. Amini , “A survey on federated learning for resource - constrained IoT devices ,” B. Yin, H. Yin, Y. Wu, and Z. Jiang , “FDC: A secure federated deep learning mechanism for data collaborations in the Internet of Things ,” J . Zhang et al., “ MobiPose : Real-time multi- person pose estimation on mobile devices ,”.

Authors. Cheng Zhang. Cheng Zhang.

Summary. “. ”. Thank you.