MetaFinger: Fingerprinting the Deep Neural Networks with Meta-training

Authors: Kang Yang, Run Wang, Lina Wang

IJCAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments show that our method achieves 99.34% and 97.69% query accuracy on average, surpassing existing methods over 30%, 25% on CIFAR-10 and Tiny-Image Net, respectively.
Researcher Affiliation Academia Kang Yang1,2 , Run Wang1,2 , Lina Wang1,2,3 1School of Cyber Science and Engineering, Wuhan University, China 2Key Laboratory of Aerospace Information Security and Trusted Computing, Ministry of Education, China 3Zhengzhou Xinda Institute of Advanced Technology
Pseudocode Yes Algorithm 1 Meta-training
Open Source Code Yes Our code is available at https://github.com/kangyang WHU/Meta Finger/
Open Datasets Yes on CIFAR-10 and Tiny-Image Net benchmark datasets.
Dataset Splits Yes We first split the meta-data into train data Dtrain and validation data Dval.
Hardware Specification No The paper does not provide specific details about the hardware used for experiments, such as GPU models, CPU specifications, or memory.
Software Dependencies No The paper does not list specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions, or other libraries).
Experiment Setup Yes In Section 5.1, the paper describes input modification attacks including 'Random Resize and Padding (RP)', 'Input Noising' (Gaussian noise and universal noise), and 'Input Smoothing' (mean, median, Gaussian kernel). It also details model modification attacks with 'Fine-tuning' (FTLL, FTAL, RTLL, RTAL modes), 'Weight pruning' (p% from 10% to 70%), and 'Weight Noising' (with α values). Algorithm 1 also lists 'Learning Rate η, Train Epoch Nepoch, Loss Control λ, Number of Sample k' as inputs for meta-training.