CamoPatch: An Evolutionary Strategy for Generating Camoflauged Adversarial Patches

Authors: Phoenix Williams, Ke Li

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we empirically evaluate our proposed method s effectiveness by attacking classifiers trained on the Image Net dataset [16]. The experimental setup is outlined in Section 4.1, followed by a comparative analysis with state-of-the-art adversarial patch methods, including Patch-RS [15], TPA [66], OPA [20], Adv-Watermark [27] and a black-box adaptation of LOAP [47] in Section 4.2. Last but not the least, Section 4.3 offers an ablation study that scrutinizes the significance of various components and parameters within our proposed method.
Researcher Affiliation Academia Phoenix Neale Williams Department of Computer Science University of Exeter Exeter, EX4 4RN pw384@exeter.ac.uk Ke Li Department of Computer Science University of Exeter Exeter, EX4 4RN k.li@exeter.ac.uk
Pseudocode Yes Algorithm 1: Evolutionary Strategy for Generating Disguised Adversarial Patches (Camo Patch)
Open Source Code No The paper does not contain an explicit statement about releasing its source code or a link to a code repository for the methodology described.
Open Datasets Yes For our experiments, we follow a similar setup to preceding works, conducting non-targeted and targeted attacks on DNN classifiers trained on the Image Net dataset [16].
Dataset Splits Yes A subset of 1, 000 images, correctly classified by each classifier from the Image Net validation set, is chosen and resized to dimensions of 224 224 3.
Hardware Specification Yes All experiments were carried out on a system with an NVIDIA Ge Force RTX 2080Ti GPU.
Software Dependencies No The paper mentions software like 'Py Torch library [44]' and 'Robust Bench library [14]' but does not provide specific version numbers for these software dependencies.
Experiment Setup Yes Parameter Settings: To select the value of ϵ, we follow the approach of Croce et al., setting ϵ = 1600. This corresponds to a patch size of 40 40, which constitutes roughly 3.2% of the total pixel count. We assign a budget of 10, 000 queries for each attack. As discussed in Section 3, our proposed method entails four free parameters: σ, lit, t, and N. For these parameters, we set σ = 0.1, t = 300, lit = 4, and N = 100.