Paper 2026/1487

LightShark: Actively Secure Machine-Learning Inference Based on Lightweight Authenticated Distributed Comparison Function

Abstract

Recently, Shark (S\&P'25) considered the problem of actively two-party secure machine learning inference using an authenticated distributed comparison function (DCF). This is the state-of-the-art work in this setting. On the other hand, Grotto (CCS'23) built a variant DCF with the key size half that of classic DCF. Unfortunately, as Shark states, \textit{it is not known how to extend Grotto to the malicious setting}. In this paper, we present the first actively secure Grotto-style DCF scheme. Our authenticated DCF is deliberately designed on the correlated GGM tree and maintains the key-size advantage of semi-honest Grotto. We further implement an actively secure ML inference framework, named LightShark, which supports efficient primitives (e.g., ReLU, spline, and truncation) and ML models (e.g., VGG-16, GPT, BERT). Compared with Shark, our LightShark outperforms by $1.49 \times \sim 2.69\times$ and reduces communication costs by $66.7\%$ for Bert-base inference. Surprisingly, for larger LLM models, the experimental evaluation demonstrates that our framework works vastly well.