Trustworthy Dataset Proof: Certifying the Authentic Use of Dataset in Training Models for Enhanced Trust

In the realm of deep learning, the veracity and integrity of the training data are pivotal for constructing reliable and transparent models. This study introduces the concept of Trustworthy Dataset Proof (TDP), which tackles the significant challenge of verifying the authenticity of training data as declared by trainers. Existing dataset provenance methods, which primarily aim at ownership verification rather than trust enhancement, often face challenges with usability and integrity. For instance, excessive operational demands and the inability to effectively verify dataset authenticity hinder their practical application. To address these shortcomings, we propose a novel technique termed Data Probe, which diverges from traditional watermarking by utilizing subtle variations in model output distributions to confirm the presence of a specific and small subset of training data. This model-agnostic approach improves usability by minimizing the intervention during the training process and ensures dataset integrity via a mechanism that only permits probe detection when the entire claimed dataset is utilized in training. Our study conducts extensive evaluations to demonstrate the effectiveness of the proposed data-probe-based TDP framework, marking a significant step toward achieving transparency and trustworthiness in the use of training data in deep learning.