Trustworthy Pre-processing of Sensor Data in Data On-Chaining Workflows for Blockchain-Based IoT Applications

Advancing Blockchain-based Federated Learning through Verifiable Off-chain Computations

“…In our future work, we want to seize on aspects discussed in Section VII-B by further investigating mechanisms to decrease the costs for verifying federated learning results, e.g., by offchaining parts of the necessary storage, and to remove yet existing trust assumptions of the learning nodes, e.g., through trustworthy pre-processing [30]. In addition, we want to take a look at approaches to scale the solution to handle larger problem sizes.…”

“…6: Accuracy scores for different numbers of participants and various batch sizes. (10,20,30,40) with a total number of 300 training cycles each. Different input data batches are also randomly chosen from the data set to avoid over-fitting of the learning model.…”

“…While for the former case no practical solution exists that does not add further trust assumptions, for the latter case endto-end integrity can be provided by applying the trustworthy pre-processing model presented in [30]. Trustworthy preprocessing enables a smart contract to verify authenticity and integrity of data that has been signed by a trusted data source but has been pre-processed by a distrusted intermediary, e.g., an oracle [31] or the off-chain learning node.…”

“…Trustworthy preprocessing enables a smart contract to verify authenticity and integrity of data that has been signed by a trusted data source but has been pre-processed by a distrusted intermediary, e.g., an oracle [31] or the off-chain learning node. In [30], it has been shown that the pre-processing model can, among others, be realized by applying zkSNARKs. Consequently, it represents a viable extension to our system to remove trust assumptions from the learning node and to enable end-to-endintegrity between the data source and the on-chain aggregator.…”

“…Compared with ZKPs, TEEs have weaker security guarantees and rely on a trusted third party, i.e., the hardware manufacturer, which is not required for ZKPs. However, they provide better performance and have a smaller computational overhead as, for example, shown in [30].…”