TDDBench: A Benchmark for Training data detection
Zhihao Zhu, Yi Yang, Defu Lian
TL;DR
TDDBench provides a comprehensive benchmark for evaluating training data detection methods across multiple datasets and modalities, highlighting current limitations and guiding future improvements in the field.
Contribution
This work introduces TDDBench, the first extensive benchmark for TDD methods, covering diverse datasets, detection paradigms, and performance metrics, facilitating systematic evaluation and comparison.
Findings
TDD algorithms generally perform poorly across datasets.
Benchmark reveals significant room for improvement in TDD effectiveness.
Open-source release promotes reproducibility and further research.
Abstract
Training Data Detection (TDD) is a task aimed at determining whether a specific data instance is used to train a machine learning model. In the computer security literature, TDD is also referred to as Membership Inference Attack (MIA). Given its potential to assess the risks of training data breaches, ensure copyright authentication, and verify model unlearning, TDD has garnered significant attention in recent years, leading to the development of numerous methods. Despite these advancements, there is no comprehensive benchmark to thoroughly evaluate the effectiveness of TDD methods. In this work, we introduce TDDBench, which consists of 13 datasets spanning three data modalities: image, tabular, and text. We benchmark 21 different TDD methods across four detection paradigms and evaluate their performance from five perspectives: average detection performance, best detection performance,…
Peer Reviews
Decision·ICLR 2025 Poster
This work provides a comprehensive evaluation of training data detection (TDD). The evaluation spans multiple data modalities, including image, tabular, and text, and implements 21 state-of-the-art TDD algorithms, including large language models such as Pythia-12B. This is the largest benchmark for TDD to date, offering a valuable reference for future research on this problem.
Although this evaluation extensively covers multiple data modalities and four main types of TDD methods, the insights from the evaluation are limited. It lacks in-depth analysis of the results. For example, in Table 4, the authors do not explain why performance on CIFAR-100 is noticeably higher than on other datasets, and similarly in Table 5, there is no explanation for why the target model CatBoost outperforms others on tabular data. Some explanations of the results are also unclear. For insta
1. The paper analyzes the latest methods' limitations and strengths carefully. Particularly, the correlation between train-test performance gap and TDD performance is interesting.
1. The proposed benchmark is not truly large-scale as it considers regular-sized datasets. 2. The proposed benchmark is not significantly different from the existing ones except that it covers different modalities. It would be more interesting if the paper does not follow data split rules from the existing benchmarks but explore something completely new. Another missing aspect is the size of data points that are evaluated under TDD. 3. The claim that TDD outperformance comes at a cost of compu
The benchmark is comprehensive. The TDDBench covers multiple data modalities and includes a wide range of datasets, algorithms, and model architectures, making it one of the most inclusive TDD benchmarks available. The writing and formatting are decent.
It will be great if the authors can propose a novel method to improve the performance based on the analysis in their benchmark experiments. It lacks an important discussion about the TDD and different training methods. There are supervised learning, self-supervised learning, unsupervised learning. Would TDD method generalizable to different training methods?
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsIntelligent Tutoring Systems and Adaptive Learning · Advanced Data Processing Techniques
MethodsSoftmax · Attention Is All You Need