Synthetic Artifact Auditing: Tracing LLM-Generated Synthetic Data Usage in Downstream Applications

TL;DR

This work defines synthetic artifact auditing as a binary task to detect whether downstream artifacts are trained on or derive from LLM-generated synthetic data. It introduces three methods—metric-based auditing, tuning-based auditing, and classification-based auditing—that operate without disclosing proprietary training details, and applies them to classifiers, generators, and statistical plots. Across three text tasks, two summarization tasks, and two visualization tasks, the framework achieves strong auditing performance, with tuning-based auditing offering the best accuracy when white-box access is available. The study demonstrates that synthetic data imprint detectable patterns in downstream artifacts, enabling practical auditing to promote transparency, accountability, and compliance in synthetic data use.

Abstract

Large language models (LLMs) have facilitated the generation of high-quality, cost-effective synthetic data for developing downstream models and conducting statistical analyses in various domains. However, the increased reliance on synthetic data may pose potential negative impacts. Numerous studies have demonstrated that LLM-generated synthetic data can perpetuate and even amplify societal biases and stereotypes, and produce erroneous outputs known as ``hallucinations'' that deviate from factual knowledge. In this paper, we aim to audit artifacts, such as classifiers, generators, or statistical plots, to identify those trained on or derived from synthetic data and raise user awareness, thereby reducing unexpected consequences and risks in downstream applications. To this end, we take the first step to introduce synthetic artifact auditing to assess whether a given artifact is derived from LLM-generated synthetic data. We then propose an auditing framework with three methods including metric-based auditing, tuning-based auditing, and classification-based auditing. These methods operate without requiring the artifact owner to disclose proprietary training details. We evaluate our auditing framework on three text classification tasks, two text summarization tasks, and two data visualization tasks across three training scenarios. Our evaluation demonstrates the effectiveness of all proposed auditing methods across all these tasks. For instance, black-box metric-based auditing can achieve an average accuracy of $0.868 \pm 0.071$ for auditing classifiers and $0.880 \pm 0.052$ for auditing generators using only 200 random queries across three scenarios. We hope our research will enhance model transparency and regulatory compliance, ensuring the ethical and responsible use of synthetic data.

Figures (37)

• Figure 1: Overview of the synthetic artifact auditing. The auditing targets are: classifiers (\ref{['section:classifier_audit']}), generators (\ref{['section:generator_audit']}), and statistical plots (\ref{['section:plot_audit']}).
• Figure 2: Overview of the metric-based auditing.
• Figure 3: Overview of the tuning-based auditing.
• Figure 4: Overview of target/reference classifier setup.
• Figure 5: Metric-based auditing performance for target classifiers fine-tuned on pre-trained DistilBERT with varying query budgets of $\mathcal{Q}_{\textit{syn}}$$\{10, 20, 50, 100, 200\}$ for $\mathcal{T}_{\mathcal{C}_3}$ in (a) $\mathcal{S}_1$ and (b) $\mathcal{S}_2$. The source LLM is GPT-3.5. The auditing performance of other source LLMs and tasks is shown in \ref{['appendix:metric_based_audit_classifier_qs_sup']}.