🤖 machine learning

Can Blockchains Reliably Train Machine Learning Models?

This paper introduces Proof of Training (PoT), a protocol that repurposes energy-intensive proof-of-work mining resources into verifiable machine learning model training, demonstrating through theoretical analysis and implementation that this approach achieves high throughput, robustness, and enhanced network security while addressing the technical gaps in decentralized learning.

Original authors: Peihao Li, Nadia Dahmani

Published 2026-05-06

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Peihao Li, Nadia Dahmani

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a massive, global competition where thousands of computers are constantly trying to solve incredibly difficult math puzzles. This is how Bitcoin and similar blockchains currently work (a system called "Proof of Work"). The computers that solve these puzzles first get a cash reward. However, there's a catch: these puzzles are useless. They don't create anything of value; they just burn a tremendous amount of electricity to keep the network secure. It's like having a fleet of race cars driving in circles just to prove they can drive fast, consuming fuel but going nowhere.

At the same time, the world of Artificial Intelligence (AI) is starving for computing power. Training smart AI models requires massive amounts of calculation, which is expensive and hard to get.

This paper proposes a solution called Proof of Training (PoT). The idea is simple: instead of wasting energy on useless math puzzles, let's use that same massive computing power to train AI models. But, how do we make sure people are actually doing the work and not cheating? That's where the paper's new protocol comes in.

Here is how the system works, broken down into everyday concepts:

1. The New Game: Training Instead of Puzzles

Think of the blockchain not as a puzzle solver, but as a global classroom.

The Client (The Teacher): Someone who needs an AI model trained posts a "homework assignment" (a dataset and a goal) on the network.
The Miners (The Students): Instead of solving puzzles, these computers download the assignment and start training the AI. They promise, "I will finish this homework by this time," and lock in their promise.
The Validators (The Proctors): These are independent computers that check the homework. They look at the finished AI model and grade it against a secret test set to see how well it performs.

2. The "Commit and Reveal" Trick (Preventing Cheating)

In a normal classroom, a student might wait to see what the test questions are before studying. In this system, that's impossible.

The Lockbox: Before the test questions (the validation data) are released, the "Student" (Miner) must put their finished AI model into a digital lockbox. They send a receipt (a hash) to the network proving they have a specific model ready.
The Reveal: Only after the test questions are released does the Miner open the lockbox and show their model.
Why this matters: This stops a cheater from looking at the test answers, solving the problem, and then pretending they solved it earlier. They are locked in before they know the answers.

3. The "Watchdog" System (Disputes)

You might ask, "Who checks the Proctors? What if they grade unfairly or are lazy?"
The paper introduces a Dispute Mechanism, similar to a "Whistleblower" system:

Assume Good Faith: The system assumes the Proctors (Validators) are honest initially to save time and energy.
The Challenge Window: For a short period after grading, anyone in the network can act as a Watchdog. If a Watchdog sees a Proctor gave a bad grade or a fake score, they can challenge it.
The Stakes: To be a Proctor, you must put up a deposit (money). If you are caught cheating or grading wrong, you lose your deposit. The Watchdog who catches you gets a cut of that money as a reward.
The Result: This creates a powerful incentive for Proctors to be honest and for Watchdogs to keep an eye on them. It's like a game where it's more profitable to be honest than to cheat.

4. Why This is Better Than Old Ideas

The paper argues that previous attempts to do this failed because they were too slow or insecure.

The Bottleneck: Old ideas tried to check every single homework assignment on the main blockchain. This would clog the system like a traffic jam.
The Solution: PoT uses a Layer 2 approach. Think of the main blockchain (Layer 1) as the Bank Vault. It only holds the final results and the money. The actual work (training and grading) happens off-site in a fast, parallel network (Layer 2). The Bank Vault only gets involved if there is a fight (a dispute) or to pay out the winners. This keeps the system fast and scalable.

5. The Security Guarantee

The authors did the math to prove this works. They showed that if the penalty for cheating (losing your deposit) is high enough, and the chance of getting caught is high enough, then honesty is the only logical choice.

If a Miner tries to fake a good model, they will get caught by a Watchdog and lose money.
If a Validator tries to give a fake grade, they will lose their deposit.
Therefore, the system naturally pushes everyone to do the real work and grade it fairly.

Summary

The paper proposes turning the "useless" energy of Bitcoin-like networks into a decentralized AI factory.

Old Way: Computers burn electricity to solve meaningless riddles.
New Way (PoT): Computers use that same electricity to train useful AI models.
How it stays honest: A system of deposits, secret test questions, and a "whistleblower" reward system ensures that no one can cheat without losing their money.

The result is a network that is fast, secure, and capable of handling many AI training jobs at once, all without wasting energy on tasks that produce nothing of value.

Technical Summary: Can Blockchains Reliably Train Machine Learning Models?

Problem Statement
Current Proof of Work (PoW) blockchains consume vast amounts of electricity solely to achieve trustless consensus, with no additional practical utility. While repurposing this computational power for Machine Learning (ML) training offers a potential solution to energy waste, existing attempts to integrate ML into blockchain consensus face significant technical gaps. Previous proposals, such as Proof of Useful Work (PoUW), Proof of Deep Learning (PoDL), and Robust Proof of Learning (RPoL), suffer from critical flaws in scalability, security, and adaptability. Specifically, many rely on storing test data on-chain (exhausting storage), require globally synchronized timestamps (fragile under network delays), assume trusted platforms, or fail to scale to modern non-deterministic GPU training. Furthermore, distributed learning protocols without robust security allow adversaries to fake rewards, breaking fairness and discouraging participation. The core challenge is designing a protocol that directs mining power toward verifiable ML training while maintaining the security and scalability guarantees of a permissionless blockchain.

Methodology
The authors propose Proof of Training (PoT), a protocol designed to bridge the gap between blockchain consensus and decentralized ML training. The methodology involves three primary components:

Theoretical Foundation & Consensus Design:
- The paper first proves a theoretical impossibility result (Theorem II.3): a permissionless system where validators are admitted without scarce resources (unresourced permissionless validators) cannot simultaneously achieve safety and liveness.
- Consequently, PoT adopts a hybrid architecture separating the consensus layer from the training layer. It utilizes a Proof of Stake (PoS) mechanism for the consensus layer (Aggregators and Validators) to ensure security, while the training layer remains open.
- The system employs an L1–L2 architecture:
  - L1 (Settlement Layer): Anchors stake, escrow, and reward distribution on a mainstream chain.
  - L2 (Execution Layer/DTN): A Decentralized Training Network (DTN) that handles the heavy lifting of task lifecycle, model claiming, validation, and dispute handling off-chain.
Protocol Mechanics (The Task Cycle):
- Participants: Clients post tasks; Miners (Service Providers) train models; Validators evaluate models; Aggregators maintain the ledger; Verifiers (any participant) challenge incorrect validations.
- Commit-and-Reveal: Miners commit to a model hash before training data is revealed, preventing model theft.
- Deterministic Scoring: Validators evaluate revealed models against a client-provided test set using a deterministic scoring function ( $VRF_{Model}$ ).
- Dispute Mechanism: Instead of validating every model on-chain (which creates bottlenecks), PoT assumes validators are honest initially. A challenge window allows any participant to dispute a validation report. If a challenge succeeds, the dishonest validator is slashed (losing a fraction of their stake), and the challenger is rewarded. This keeps on-chain state lightweight.
Incentive Security Analysis:
- The authors model the learning layer as a sequential game. They derive sufficient conditions (Theorem II.9) under which truthful validation is a strictly dominant strategy for validators and maximizing true model score is the best response for miners.
- Key conditions include: the probability of detection ( $q$ ) multiplied by the slashing fraction ( $\phi$ ) and stake ( $s$ ) must exceed the external gain from cheating plus the cost difference between honest and dishonest reporting.

Key Contributions

Proof of Training (PoT) Protocol: A novel consensus mechanism that redirects PoW-style computing resources toward verifiable ML training.
Theoretical Security Proof: A formal proof demonstrating that unresourced permissionless validators cannot secure a ledger, justifying the shift to a stake-weighted consensus layer for PoT.
Scalable Architecture: An L1–L2 design that decouples value settlement from utility execution, allowing heavy training and validation to occur off-chain while settling only compact outcomes on-chain.
Incentive Compatibility: A game-theoretic analysis proving that the protocol's economic parameters (slashing, rewards, challenge costs) can align participant incentives with honest behavior, ensuring reliable model training.

Results

Theoretical Validation: The authors demonstrate that under specific parameter settings (slashing fraction, detection probability, and stake), honest behavior is the strictly dominant strategy for validators, and rational challengers will sustain monitoring.
Throughput Evaluation: Simulations of the DTN using PBFT-style workflows with committee sizes of 10, 30, and 50 nodes show that the system can achieve high throughput.
- Under medium to fast bandwidth conditions, synchronization times remain in the range of seconds even as batch sizes grow.
- The system is shown to comfortably meet the scalability requirements of reported industrial deep learning job arrival rates (typically well below one job per second).
Security Simulations:
- Exp1: Visualizes the payoff gap, confirming that truthful validation becomes strictly dominant when the slashing risk exceeds the cost of cheating.
- Exp2: Demonstrates that monitoring can be sustained in equilibrium when expected slashing rewards cover verification costs.
- Exp3: Shows that high detectability (driven by active monitoring) incentivizes miners to invest effort in improving true model scores, whereas low detectability leads to reduced effort and lower quality.

Significance and Claims
The paper claims that PoT provides a practical foundation for scalable and adversarially robust open training networks. Its significance lies in:

Energy Efficiency: It offers a pathway to repurpose the massive computational power of PoW networks for valuable ML tasks, addressing the environmental concerns of blockchain.
Reliability: By combining a stake-weighted consensus layer with a dispute-driven verification mechanism, it ensures that training outputs are reliable and rewards are allocated fairly without requiring every node to re-run expensive training tasks.
Scalability: The L1–L2 separation allows the network to handle a broad range of workloads, including large model training and privacy-sensitive collaborative training, without congesting the ledger.
Adaptability: The design supports frequent updates to training workloads and evaluation rules off-chain, avoiding the rigidity and coordination costs associated with on-chain consensus updates.

The authors conclude that PoT successfully bridges the technical gaps of previous proposals, offering a system that is secure, scalable, and capable of supporting a decentralized market for machine learning training.