Data-Free Universal Attack by Exploiting the Intrinsic Vulnerability of Deep Models

Authors: YangTian Yan, Jinyu Tian

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Remarkably, our method achieves highly competitive performance in attacking popular image classification deep models without using any image samples. We also evaluate the black-box attack performance of our method, showing that it matches the state-of-the-art baseline for data-free methods on models that conform to our theoretical framework. Experiments demonstrate that the attack success rate decreases by only 4% when the adversary has access to just 50% of the linear layers in the victim model.
Researcher Affiliation Academia Yang Tian Yan, Jinyu Tian* Faculty of Innovation Engineering, Macau University of Science and Technology EMAIL, EMAIL
Pseudocode Yes Algorithm 1: Algorithm for Our Proposed Intri UAP Method
Open Source Code Yes Code https://github.com/yyt0718/Intri Attack
Open Datasets Yes We benchmarked our method against the latest datafree and data-dependent UAPs on the Image Net dataset. ... To effectively assess the attack performance of our method, we report the fooling ratio on the 50,000-image validation set from Image Net ILSVRC2012
Dataset Splits Yes To effectively assess the attack performance of our method, we report the fooling ratio on the 50,000-image validation set from Image Net ILSVRC2012
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU models, CPU types, or memory specifications) used for running the experiments.
Software Dependencies No The paper does not specify any software dependencies with version numbers (e.g., specific libraries or programming language versions).
Experiment Setup Yes Initialization of Intri UAP. We consider the following initialization of our Intri UAP ξ describe in Algorithm 1: 1. Image Net Mean and Range prior... 2. Gaussian Distribution: We generated perturbations by sampling from a Gaussian distribution N(µ, σ2). In our experiments, µ was set to 0.45, with σ values of 0.1. 3. Uniform Distribution: We generated perturbations by sampling from a uniform distribution U(a, b). In our experiments, a was set to 0.40, b was set to 0.60. ... Finally, a clipping operation is applied to ensure that the ℓ -norm of ξ remains within the bound of 10. ... We employ the Adam optimizer combined with a Step LR scheduler.