Bionetta: Efficient Client-Side Zero-Knowledge Machine Learning Proving

TL;DR

Bionetta addresses client-side zero-knowledge proving for neural networks by integrating an R1CS-based Groth16 framework with UltraGroth lookup tables and circuit-embedded weights, enabling succinct proofs and efficient on-chain verification ($|\pi|$ constant and four pairings). It combines float quantization with a ReLU-based activation strategy and a quantization/overflow management scheme to map real-valued inference to finite-field arithmetic, while preserving accuracy within tight bounds. The key contributions are (i) circuit-embedded weights that drastically reduce witness size, (ii) a quantization/overflow framework with provable error bounds, (iii) UltraGroth enabling lookup-based non-linear-cost reduction, and (iv) extensive experiments showing substantial improvements over state-of-the-art zkML frameworks, including mobile-device proving. The practical impact is enabling deployable, privacy-preserving ML proofs on Ethereum-compatible runtimes and edge devices, with an emphasis on client-side proving and compact verification, though the approach remains non-universal due to public-weights constraints.

Abstract

In this report, we compare the performance of our UltraGroth-based zero-knowledge machine learning framework Bionetta to other tools of similar purpose such as EZKL, Lagrange's deep-prove, or zkml. The results show a significant boost in the proving time for custom-crafted neural networks: they can be proven even on mobile devices, enabling numerous client-side proving applications. While our scheme increases the cost of one-time preprocessing steps, such as circuit compilation and generating trusted setup, our approach is, to the best of our knowledge, the only one that is deployable on the native EVM smart contracts without overwhelming proof size and verification overheads.

Figures (9)

• Figure 1: Relative resources overhead of existing zkML frameworks compared to Bionetta (UltraGroth). Lower values are better.
• Figure 2: Architecture of the Bionetta framework
• Figure 3: The precision cut operation in the Bionetta system. The dashed magenta boxes represent the layers of the neural network. The dashed blue box represents the output of the layer $i$, which is cut by $\ell_i\rho$ bits. Finally, green arrows represent the pre- and post-processing of the data done outside the circuit.
• Figure 4: The witness includes three parts: the public part $\mathcal{I}_X$, the private part $\mathcal{I}_W$ split into $d+1$ rounds $\{\textcolor{blue!50!black}{\mathcal{I}_W^{\langle i \rangle}}\}_{i \in [d+1]}$ and the sampled challenges $\{\mathcal{I}_R^{\langle i \rangle}\}_{i \in [d]}$.
• Figure 5: The prover complexity depending on the chunk size $w$. The different colors represent different values of $L$.

Theorems & Definitions (25)

• Example
• Example
• Proposition
• Definition
• Proposition
• Definition
• Example
• Theorem
• Example
• Proposition