Keskeiset käsitteet
A novel defense strategy, COIN, is proposed to effectively mitigate the impact of convolution-based unlearnable datasets by employing random pixel-based image transformations.
Tiivistelmä
The paper focuses on addressing the challenge of defending against a new type of unlearnable datasets (UDs) called convolution-based UDs, which have been shown to render existing defense mechanisms ineffective.
Key highlights:
The authors first model the convolution-based UDs as the result of multiplying a matrix by clean samples, and propose two metrics, Θimi and Θimc, to quantify the inconsistency within intra-class multiplicative noise and the consistency within inter-class multiplicative noise, respectively.
Validation experiments show that increasing both Θimi and Θimc can mitigate the unlearnable effect of convolution-based UDs.
The authors then design a random matrix transformation, Ar, to boost both Θimi and Θimc, and extend this idea to propose a new defense strategy called COIN, which employs random pixel-based image transformations via bilinear interpolation.
Extensive experiments demonstrate that COIN significantly outperforms state-of-the-art defenses against existing convolution-based UDs, achieving an improvement of 19.17%-44.63% in average test accuracy on the CIFAR-10 and CIFAR-100 datasets.
The authors also propose two new types of convolution-based UDs, VUDA and HUDA, and show that COIN is the most effective defense against them.
Tilastot
The test accuracy of models trained on CUDA UD without defense is around 20-27%.
Applying AT and JPEG compression can improve the test accuracy to around 36-42%.
Our proposed defense COIN can achieve a test accuracy of 61.35% on average, outperforming existing defenses by 19.17%-44.63%.
Lainaukset
"To the best of our knowledge, none of the existing defense mechanisms demonstrate efficacy in effectively mitigating convolution-based UDs."
"Extensive experiments reveal that our approach significantly overwhelms existing defense schemes, ranging from 19.17%-44.63% in test accuracy on CIFAR-10 and CIFAR-100."