# PoLO: Proof-of-Learning and Proof-of-Ownership at Once with Chained Watermarking

Source: [PoLO: Proof-of-Learning and Proof-of-Ownership at Once with Chained Watermarking](<https://www.sota2.com/research/paper/68fa789a1e209131cea793d4>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-04-28T05:00:04.744Z

Authors: Haiyu Deng, Yanna Jiang, Guangsheng Yu, Qin Wang, Xu Wang, Baihe Ma, Wei Ni, Ren Ping Liu

Published: 2025-05-18T08:19:18.000Z

## Abstract

Our evaluation shows that PoLO achieves \\textbf{99\\%} watermark detection accuracy for ownership verification, while preserving data privacy and cutting verification costs to just \\textbf{1.5--10\\%} of traditional methods. Forging PoLO demands \\textbf{1.1--4$\\times$} more resources than honest proof generation, with the original proof retaining over \\textbf{90\\%} detection accuracy even after attacks.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [LLM alignment on WildChat (train)](<https://www.sota2.com/research/sota/llm-alignment-on-wildchat-train.md>): LLM Alignment; WildChat (train); Loss: 1.1951.

- [LLM Alignment on WikiText-2 raw v1 (train)](<https://www.sota2.com/research/sota/llm-alignment-on-wikitext-2-raw-v1-train.md>): LLM Alignment; WikiText-2 raw v1 (train); Loss: 3.5737.

- [Image Classification on CIFAR-10 (Proof Generation and Verification Time)](<https://www.sota2.com/research/sota/image-classification-on-cifar-10-proof-generation-and-verification-time.md>): Image Classification; CIFAR-10; Training + Proof Generation Time (s): 990.74.

- [Image Classification on CIFAR-100 (Training and Proof Generation Timings)](<https://www.sota2.com/research/sota/image-classification-on-cifar-100-training-and-proof-generation-timings.md>): Image Classification; CIFAR-100; Training Time (s): not recorded.

- [Image Classification on TinyImageNet (Proof Overhead Timing)](<https://www.sota2.com/research/sota/image-classification-on-tinyimagenet-proof-overhead-timing.md>): Image Classification; TinyImageNet; Training Time (s): not recorded.

- [Text Classification on AG News (Training and Proof Time)](<https://www.sota2.com/research/sota/text-classification-on-ag-news-training-and-proof-time.md>): Text Classification; AG-News; Training Time (s): not recorded.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2505.12296>)

[PDF](<https://arxiv.org/pdf/2505.12296.pdf>)