DICE: Decentralized Incentive-driven Federated Learning for Efficient & Scalable Customization of Pretrained Models in Web3

What it claims

DICE targets one bottleneck in blockchain-powered federated learning: every validator re-verifies every worker update each round, so per-validator cost grows linearly with the number of workers. Its Tree-structured Committee Consensus (TCC) splits verification and aggregation into tree-structured sub-tasks and sends only some of them to committee voting. The authors claim this makes per-validator workload grow sub-linearly with worker count, while bounding the success probability of malicious behaviour. The paper is accepted at CIKM '26 (Rome, November 2026), runs 27 pages, and is licensed CC BY 4.0.

Why existing BCFL breaks on large models

Table 1 measures BERT-Large (24 layers, 340M parameters) under prior BCFL algorithms. The setting is an A100, FP32, Adam, sequence length 512, and the 2018 n2c2 NER dataset.

• Worker memory grows with batch size: 11 GB at batch 8, 17 GB at 16, 30 GB at 32 and 59 GB at 64.
• Validator memory for concurrent verification of all updates grows with worker count: 18 GB at 10 workers, 71 GB at 40, 125 GB at 70 and 170 GB at 100.

The authors argue that PEFT methods (LoRA, BitFit) cut only worker cost. Distillation, pruning and quantization cut validator cost but may hurt model quality. None of these removes the linear validator bottleneck.

How it works

Three roles take part. A requester funds a smart contract holding the task spec, validation data and cryptocurrency rewards. Workers train on private data. Validators verify and aggregate updates. Model parameters live on IPFS, with hashes and file handles stored on-chain. A participant whose file is unreachable, or whose hash does not match, gets no reward.

Each round has four stages:

1. Preparation. Workers and validators who cannot hold the full model get a model slice. The layers outside their slice are compressed by knowledge distillation into an "emulator" that the requester tunes.
2. Training. Dynamic Progressive Learning trains part of the hidden layers, around frozen emulator layers.
3. VA_COMMIT. Validators submit hashed fingerprints and commitments of their results. This stops fraud by copying, and DICE modifies the step slightly to speed up TCC.
4. VA_REVEAL. Validators publish the raw results. Only sub-tasks with conflicting results go to a committee vote.

Model Slice Verification is scoring-based: it uses thresholds on validation performance and stays compatible with zero-knowledge proofs. Eligibility to validate comes from economic staking. The Efficiency-Driven Incentive scheme penalizes validators who create conflicts, because each conflict forces a vote and erodes TCC's savings.

Non-responsive participants are handled by a per-state timeout set from a moving average of that state's execution time. After a timeout, any participant can end the task, and the requester then restarts it or keeps the model trained so far.

Threat model

The scheme defends against workers and validators who poison, free-ride by copying others' results, or stop responding. These attackers may collude and may use Sybil or DDoS attacks. The blockchain and IPFS are assumed reliable. Requesters are assumed rational and honest. Delivery delays may be unbounded, but repeated messages are eventually delivered.

Limits

Privacy inference attacks, and defences such as differential privacy, are explicitly out of scope. Liveness assumes at least one honest participant stays active per task. The fetched text stops in the preliminaries, so DICE's own experimental numbers are not shown. That includes its cost reductions, accuracy against baselines and attack-resistance results. No code release is mentioned.