Skip to content
Crypto.world

The Ledger

Paper Finds ZK Rewards Raised Attackers’ Share in Federated Learning Tests

A study found that its privacy-preserving token reward design gave simulated adversaries a larger share of rewards than a version that disclosed quality scores. The paper also reports late-stage adversary monopolies and higher transaction costs for the zero-knowledge design.

Paper Finds ZK Rewards Raised Attackers’ Share in Federated Learning Tests

Privacy gate, larger adversary share

A privacy-preserving token reward design gave simulated attackers an average 24.51% of rewards, compared with 10.27% in a version that disclosed quality scores, a paper reports. The researchers say the zero-knowledge design also produced 9.8 times greater variation in attackers’ reward share across their tests.

Source arxiv.org

Submitted to arXiv on November 11, the paper compares two Solidity smart contracts that issue non-transferable tokens for contributions to federated learning. One distributes rewards using a public quality score; the other uses a zero-knowledge proof to verify that a score meets a threshold without disclosing the score itself. The paper was accepted at the 12th Balkan Conference in Informatics (BCI 2026), where it received a Best Paper Award.

How the scoring shifted

The experiment used 30 clients to train a LeNet-5 model on MNIST over 20 rounds, sampling 10 clients per round. Five adversaries submitted zero gradients, random noise or stale updates, among other behaviors, and the researchers repeated the experiment across three random seeds. The privacy-preserving contract required a quality score of at least 50.

The authors link the results to a bias in cosine-similarity scoring. As training progressed and the global model converged, honest clients’ updates shrank, while noise updates could still score above the fixed threshold. In the seed-42 run, the noise adversary became the only verified participant in rounds 17 through 20 and captured all tokens minted in those rounds.

Across all three seeds, the zero-knowledge design produced an average adversary reward share of 24.51%, with a standard deviation of 6.93 percentage points. The corresponding figures for the public-score contract were 10.27% and 0.71 percentage points. The paper also reports that the Gini coefficient for honest participants’ rewards rose from 0.2171 under the standard contract to 0.4616 under the zero-knowledge version.

Costs and limits

The paper’s gas measurements show 13.0 million gas for the standard contract over a 20-round training run and 16.2 million for the zero-knowledge version using a stub verifier. With a production Groth16 verifier, the researchers project 29.8 million gas. Their model estimates a single reward-distribution transaction can cover 271 participants under the zero-knowledge design.

The authors caution that the evaluation used a stub verifier to isolate contract costs and assumed an honest coordinator. Secure aggregation was not used, so the coordinator could see plaintext gradients. The proof hides individual quality scores, but the paper says on-chain proof timing, reward minting and token burning remain observable. It discusses possible mitigations including leave-one-out aggregation, dynamic thresholds and tracking participants’ historical quality.

Sources

  1. Original source · Preprint Tokenized Reward Systems for Federated Learning: The Privacy vs. Fairness Tradeoff arXiv:2611.16065

Related stories

Explore

More articles