Proof of Reasoning for Privacy Enhanced Federated Blockchain Learning at the Edge

TL;DR

This paper tackles privacy concerns in blockchain-enabled federated learning for IoT and healthcare by introducing Proof of Reasoning (PoR), a three-stage, privacy-preserving consensus mechanism that uses a masked autoencoder (MAE) as a lightweight edge feature extractor and a downstream classifier, with encoded data shared via Encoder-Decoder Interface (EDI) transactions. PoR enables verifiable, ranking-based aggregation on the blockchain, reducing reliance on potentially malicious participants and supporting advanced aggregation schemes while maintaining low latency and storage growth. The approach is validated through CIFAR-10, transfer learning, and ChestMnist/PneumoniaMnist experiments, showing that high masking ratios can preserve downstream accuracy and that encoder-based transfer learning can boost performance in federated settings. Overall, PoR provides a practical, privacy-preserving pathway for scalable edge federated learning with blockchain-based verification, with potential extensions to multi-channel MAEs and enhanced privacy techniques.$PoR$ integrates masked encodings and verifiable aggregation to enable robust, privacy-preserving federated learning at the edge, addressing critical challenges in IoT healthcare deployments and beyond.

Abstract

Consensus mechanisms are the core of any blockchain system. However, the majority of these mechanisms do not target federated learning directly nor do they aid in the aggregation step. This paper introduces Proof of Reasoning (PoR), a novel consensus mechanism specifically designed for federated learning using blockchain, aimed at preserving data privacy, defending against malicious attacks, and enhancing the validation of participating networks. Unlike generic blockchain consensus mechanisms commonly found in the literature, PoR integrates three distinct processes tailored for federated learning. Firstly, a masked autoencoder (MAE) is trained to generate an encoder that functions as a feature map and obfuscates input data, rendering it resistant to human reconstruction and model inversion attacks. Secondly, a downstream classifier is trained at the edge, receiving input from the trained encoder. The downstream network's weights, a single encoded datapoint, the network's output and the ground truth are then added to a block for federated aggregation. Lastly, this data facilitates the aggregation of all participating networks, enabling more complex and verifiable aggregation methods than previously possible. This three-stage process results in more robust networks with significantly reduced computational complexity, maintaining high accuracy by training only the downstream classifier at the edge. PoR scales to large IoT networks with low latency and storage growth, and adapts to evolving data, regulations, and network conditions.

Figures (5)

• Figure 1: An example of PoR being used by three participants.
• Figure 2: MAE Training and upstream feature map usage. In contrast to vanilla MAE, where the input is split into unmasked patches which are then fed sequentially into a trained encoder resulting in an encoding per patch, PoR MAE continues to mask the input, as it was when training the encoder, and only the remaining (unmasked) patches are fed to the trained encoder resulting in significantly fewer encoded patches being passed to the classifier.
• Figure 3: Self-supervised training process of a masked autoencoder (MAE): The input image is divided into patches, with a high percentage masked out. The unmasked patches are assigned positional encodings and processed by the encoder, which transforms them into encoded representations. These encoded patches are then combined with masked patches, each containing a learnable mask token and its positional encoding, and passed to the decoder, which reconstructs the original image. This training enables the encoder to extract meaningful representations from limited input data.
• Figure 4: An example of generating an Encoder-Decoder Interface (EDI) transaction to be added to the blockchain using PoR. The unmasked patches of a single datapoint are encoded by the participant's trained encoder transformer and added to the transaction as the encoded data array $\kappa$. Additionally, the weights of the downstream classifier $\omega$, the output of the downstream classifier on the encoded data $\hat{y}$ and the true classification of the input $y$ are also added to the transaction.
• Figure 5: The structure of the residual bottleneck layer: The layer consists of three sequential sub-layers. First, a 1D convolution is applied to reduce the input size, with the number of features reduced to $\frac{1}{8}$ of the original. Next, the second sub-layer performs 1D convolution with a kernel size of 3, maintaining the reduced feature dimensions. Finally, the third sub-layer restores the feature dimensions to their original size using a 1D convolution, restoring the number of features to its original size. The output of this sequence is then added to the original input, creating a residual connection to enhance feature learning.