Is there such soft blockchain concept?

Question

I define a soft blockchain as a decentralized and distributed network that computes cryptography using a neural network for every peer, thus resulting in a soft hash instead of a hard hash like SHA-256.

The application might solve social science problems, especially laws that are related to plagiarism.

For example, an artist uploading art to the decentralized network, the miners (I would say the validators or peers) would be performing continuous learning forward and backpropagation using a neural network.

So, when a plagiarist uploads similar art (not identical), like just adding some noise, it will be detected as plagiarism.

In order to persist from noise, miners must forward the input with some noise while the ground truth is kept clean.

I know this concept is not defined rigorously yet. But, overall, it's similar to that.

Answer 1

By leveraging neural networks machine learning and decentralized networks principles, your idea shifts the blockchain's focus from deterministic data integrity to probabilistic similarity assessment, where the miners or validators not only validate transactions like in traditional blockchains but also participate in continuously improving shared learning models. The hard hash like SHA-256 is crucial in blockchain's security by ensuring immutability. A soft hash in your context seems to be a probabilistic adaptive cryptographic method involving feature learning, where the output isn't a fixed cryptographic fingerprint but a learned representation of the input data such as artwork. This would allow for more nuanced detection of plagiarism, as small variations in the artwork that might otherwise evade traditional cryptographic checks could still be flagged as similar by the network.

However, implementing this in a blockchain would require efficient protocols for sharing updates to the model and synchronizing training across peers to be scalable. Training neural networks for every peer in a decentralized network could be computationally expensive, so you might aggregate validators' NNs periodically (e.g., using federated learning principles) rather than training independently. Additionally in traditional blockchains mining rewards drive validation, so for your soft blockchain to work, you would need to establish a similar long-term effective reward structure (e.g., base fee, training accuracy fee, slashing and inactivity penalties, etc) that encourages peers to engage in continuous learning and accurately detect plagiarism.

Finally consensus on what's plagiarism could be complex especially when artworks are highly subjective, which might require a form of decentralized governance to decide what is flagged as plagiarism. Unlike traditional blockchain consensus mechanisms such as PoW/PoS with deterministic verification of any proposed transaction block, your soft blockchain introduces a probabilistic element to verify any proposed plagiarized or original data block, so you'll need an extremely robust mechanism to decide agreement across decentralized validators about what constitutes plagiarism or originality.