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 [1].
On Inherent Adversarial Robustness of Active Vision Systems
Authors: Amitangshu Mukherjee, Timur Ibrayev, Kaushik Roy
TMLR 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct a comprehensive robustness analysis across three categories: adversarially crafted inputs evaluated under transfer attack scenarios, natural adversarial images, and foreground-distorted images. By learning from focused, downsampled glimpses at multiple distinct fixation points, these active methods significantly enhance the robustness of passive networks, achieving a 2-21% increase in accuracy. This improvement is demonstrated against state-of-the-art transferable black-box attack. On Image Net-A, a benchmark for naturally occurring hard samples, we show how distinct predictions from multiple fixation points yield performance gains of 1.5-2 times for both CNN and Transformer based networks. |
| Researcher Affiliation | Academia | Amitangshu Mukherjee EMAIL Elmore Family School of Electrical and Computer Engineering Purdue University Timur Ibrayev EMAIL Elmore Family School of Electrical and Computer Engineering Purdue University Kaushik Roy EMAIL Elmore Family School of Electrical and Computer Engineering Purdue University |
| Pseudocode | No | The paper describes the inference processes for FALcon and GFNet using high-level overview figures (Figure 3 and Figure 4) and descriptive text. However, it does not include any clearly labeled pseudocode blocks or formal algorithms. |
| Open Source Code | Yes | 1The code is available at Git Hub. |
| Open Datasets | Yes | We perform our extensive robustness analysis on Imagenet (Deng et al., 2009), a standard benchmark for image classification. We provide similar detailed analysis for the naturally adversarial Image Net-A dataset, demonstrating 1.5-2 times improvement over the passive ventral method. |
| Dataset Splits | Yes | Adversarial samples are generated using the entire 50,000-sample Image Net test set. We follow the experimental setup outlined in the original paper (Zhang et al., 2023) and present results on a test set of 1,000 randomly selected images from the Image Net validation set (Deng et al., 2009). |
| Hardware Specification | No | The paper does not explicitly mention specific GPU models, CPU models, or detailed hardware specifications used for running the experiments. |
| Software Dependencies | No | We utilize Torchattacks (Kim, 2020), an integrated library for generating adversarial attacks (Ravikumar et al., 2022) with Py Torch, to generate adversarial samples. The paper mentions 'Py Torch' and 'Torchattacks' but does not specify their version numbers, which is required for a reproducible description of software dependencies. |
| Experiment Setup | Yes | We utilize Image Net pre-trained weights for GFNet and FALcon without any additional fine-tuning. Following the active vision structures highlighted in Section 3, we employ GFNets with Res Net50 as both global f_{G} and local f_{L} encoders. Both encoders are trained on downsampled resolution images of (96, 96) pixels. For FALcon we employ VGG16 (Simonyan & Zisserman, 2015) as the dorsal f_{D} stream, and Res Net50 as ventral f_{V}. We conduct L_{\infty } attacks with 10 iterative steps, \alpha = 2/255 , and \epsilon = 8/255 for all six iterative attacks including Auto-PGD (Croce & Hein, 2020). |