Abstract
Large-scale computational experiments, often running over weeks and over large datasets, are used extensively in fields such as epidemiology, meteorology, computational biology, and healthcare to understand phenomena and design high-stakes policies. Given that such shared results form the basis of critical inferences and policies, it is essential that the computations are validated and trusted. In particular, in a multi-agent environment involving several independent computing agents, a notion of trust in results generated by peers is critical in facilitating transparency, accountability, and collaboration. Using a novel combination of distributed validation of atomic computation blocks and a blockchain-based immutable audit mechanism, this work proposes a universal framework for distributed trust in computations. In particular, we address the scalability problem by reducing the storage and communication costs using a lossy compression scheme. This framework guarantees not only verifiability of final results but also the validity of local computations, and its cost-benefit tradeoffs are studied using a synthetic example of training a neural network.
Introduction
In multi-agent computational systems—like the OpenMalaria platform used for public health policy or distributed machine learning networks—independent agents frequently share data, models, and results. However, these agents often lack mutual trust; malicious actors might alter results, or resource-constrained agents might introduce errors.
The traditional solution of peer recomputation is prohibitively expensive for massive simulations. While blockchain technology offers immutable, verifiable records, it typically scales poorly for high-dimensional workloads. This paper introduces a "Multi-Agent Blockchain Framework" (MBF) that tackles the challenge of making blockchain-based trust scalable for heavy, iterative computational workloads by combining distributed validation with lossy compression and subsampling.
Methodology
The Multi-Agent Blockchain Framework (MBF)
The MBF functionally divides the network into three roles: Clients (who run the computations), Endorsers (who validate the steps), and Orderers (who add validated blocks to the ledger).
Scalability Techniques
To overcome the scalability bottleneck, the framework introduces several novel techniques:
- Framing and Delta Encoding: Instead of treating every iteration as a transaction, the Client groups a sequence of states into a "frame." Leveraging the fact that physical simulations and ML algorithms generally evolve smoothly (Lipschitz continuity), the Client only transmits the differences (deltas) between consecutive states.
- Lossy Compression: State updates are compressed using vector quantization. The Client reports a slightly lossy, compressed version of the computation to reduce bandwidth.
- Tolerance-Based Validation: Endorsers decompress the frame and recompute the steps. Because the compression is lossy, the recomputed state won't match the reported state perfectly. Instead, Endorsers check if the reported state falls within a strict mathematical margin ($\Delta_{marg}$).
- Subsampled Verification: Once a frame is validated, Orderers subsample the data (e.g., storing 1 out of every $K$ iterates) and store it on the blockchain with a looser verification margin, drastically reducing long-term storage requirements.
Results and Discussion
The framework was evaluated using the MNIST database, simulating a client training a 3-layer neural network using mini-batch Stochastic Gradient Descent (SGD).
- Theoretical Efficiency: The authors prove that their "Batch Mode" approach reduces communication and storage costs from scaling linearly with the raw state size to scaling logarithmically with the quantization error.
- Empirical Accuracy Boost: Surprisingly, validation acts as a regularizer. A Client restricted to a small batch size of 10 normally achieves suboptimal accuracy. However, under the strict validation requirements of the MBF, the network achieved ~96.5% accuracy—nearly matching a model trained with a much larger batch size of 50. The validation process naturally filters out extreme outliers and spurious gradients.
- Tunable Trade-offs: The system allows users to tune the validation tolerance. Stricter tolerances increase computation costs (due to rejections) but decrease communication costs and improve final model precision.
Conclusion
The MBF demonstrates that trust can be established in large-scale computations by decoupling immediate validation from long-term auditing. By leveraging the mathematical properties of simulations, the framework achieves scalability without sacrificing integrity. Future work will address network heterogeneity and more refined allocation policies for assigning Endorsers based on computational costs and latencies.
Summary updated