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 Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..

Boosting the Uniqueness of Neural Networks Fingerprints with Informative Triggers

Authors: Zhuomeng Zhang, Fangqi Li, Hanyi Wang, Shi-Lin Wang

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate that our method can be seamlessly incorporated into any existing fingerprinting scheme to facilitate the copyright tracing of deep neural networks.
Researcher Affiliation Academia Zhuomeng Zhang Shanghai Jiao Tong University EMAIL Fangqi Li Shanghai Jiao Tong University EMAIL Hanyi Wang Shanghai Jiao Tong University EMAIL Shi-Lin Wang Shanghai Jiao Tong University EMAIL
Pseudocode Yes Algorithm 1 Computing I0 tn|t1:(n 1) . Algorithm 2 Greedily selecting ˆN informative triggers.
Open Source Code Yes C Code Repo Link All codes for reproducibility in https://github.com/zzmsmm/Informative_Triggers.
Open Datasets Yes Following the settings of existing studies [29], we conducted experiments on four classification tasks: MNIST [30], Fashion MNIST [31], CIFAR-10 [32], and Image Net [33].
Dataset Splits Yes Benign [4, 37, 38]. Each trigger was randomly drawn from the training dataset.
Hardware Specification Yes We used four Ge Force RTX 2080 Ti GPUs for acceleration.
Software Dependencies No All experiments were implemented using the Py Torch framework. (No version number provided for PyTorch or other libraries).
Experiment Setup Yes The sources of heterogeneity were (I) Four network architectures including Le Net-5 [34], VGG-16 [35], Res Net-18, and Res Net-34 [36]. (II) Five learning rates ranging from 0.02 to 0.1. (III) Two learning schedules with step lengths 5 and 10. (IV) Five training epochs ranging from 10 to 60. (V) Three random downsampling of training data. So, there were P = 600 models for each task. We used four Ge Force RTX 2080 Ti GPUs for acceleration.