Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
PASS: Private Attributes Protection with Stochastic Data Substitution
Authors: Yizhuo Chen, Chun-Fu Chen, Hsiang Hsu, Shaohan Hu, Tarek F. Abdelzaher
ICML 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The comprehensive evaluation of PASS on various datasets of different modalities, including facial images, human activity sensory signals, and voice recording datasets, substantiates PASS s effectiveness and generalizability. 5. Experiments: We thoroughly evaluated PASS on three multi-attribute benchmark datasets, each representing a different application of a different modality. These datasets include Audio MNIST (Becker et al., 2018), containing recordings of human voices; Motion Sense (Malekzadeh et al., 2019), consisting of human activity sensory signals; and Celeb A (Liu et al., 2015), containing facial images. |
| Researcher Affiliation | Collaboration | Yizhuo Chen 1 2 Chun-Fu (Richard) Chen 2 Hsiang Hsu 2 Shaohan Hu 2 Tarek Abdelzaher 1 ... 1Department of Computer Science, University of Illinois Urbana-Champaign, USA 2Global Technology Applied Research, JPMorgan Chase, USA. Correspondence to: Yizhuo Chen <EMAIL>, Chun-Fu (Richard) Chen <EMAIL>. |
| Pseudocode | Yes | Algorithm 1 PASS Training Pseudo-code ... Algorithm 2 PASS Inference Pseudo-code |
| Open Source Code | No | The paper does not contain any explicit statements about releasing source code, a link to a repository, or mentions of code in supplementary materials. |
| Open Datasets | Yes | We thoroughly evaluated PASS on three multi-attribute benchmark datasets, each representing a different application of a different modality. These datasets include Audio MNIST (Becker et al., 2018), containing recordings of human voices; Motion Sense (Malekzadeh et al., 2019), consisting of human activity sensory signals; and Celeb A (Liu et al., 2015), containing facial images. |
| Dataset Splits | Yes | Audio MNIST Dataset ... The dataset contains 30,000 audio clips, divided into 24,000 for training and 6,000 for validation. ... Motion Sense Dataset ... segmented the datasets into 74,324 samples, each with a length of 128. ... Table 6. Training-testing split 7:4 ... Celeb A Dataset ... We used the official split for training and validation. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used, such as GPU models, CPU types, or memory amounts. |
| Software Dependencies | No | Table 6 lists 'Optimizer Adam W (Loshchilov & Hutter, 2019)' but does not specify the version of the underlying deep learning framework (e.g., PyTorch, TensorFlow) or any other key libraries with their version numbers. Appendix E.1 mentions 'Hu BERT-B (Hsu et al., 2021)' but this refers to a model used for feature extraction, not a general software dependency with version details. |
| Experiment Setup | Yes | Table 6. Detailed configurations of our experiments datasets, models, and optimization techniques. Optimizer Adam W (Loshchilov & Hutter, 2019) Learning rate 0.001 0.001 0.0001 Weight decay 0.0001 Learning rate scheduler Cosine scheduler Embeddings f(x) and g(x ) dimension 512 Pฮธ(X |X) training epochs 2000 200 50 Probing Attack training epochs 2000 200 50 ... Unless otherwise specified, we set ฮป = N/M and ยต = 0.2N throughout our experiments to balance private attributes protection, useful attributes preservation, and general feature preservation. The substitute dataset is constructed by randomly sampling 4096 data points from the training dataset. |