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 Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Manipulating Feature Visualizations with Gradient Slingshots
Authors: Dilyara Bareeva, Marina Höhne, Alexander Warnecke, Lukas Pirch, Klaus-Robert Müller, Konrad Rieck, Sebastian Lapuschkin, Kirill Bykov
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our approach on several DNN architectures, demonstrating its ability to replace faithful FVs with arbitrary targets. These results expose a critical vulnerability: auditors relying solely on FV may accept entirely fabricated explanations. To mitigate this risk, we propose a straightforward defense and quantitatively demonstrate its effectiveness. |
| Researcher Affiliation | Collaboration | 1Fraunhofer Heinrich Hertz Institute 2UMI Lab, ATB Potsdam 3University of Potsdam 4BIFOLD 5Machine Learning and Security Group, TU Berlin 6Machine Learning Group, TU Berlin 7Department of Artificial Intelligence, Korea University 8Max-Planck Institute for Informatics 9Centre of e Xplainable Artificial Intelligence, TU Dublin 10 Munich Center for Machine Learning (MCML) 11TU Munich |
| Pseudocode | No | The paper describes the method using mathematical formulations and descriptive text, such as in Section 3 'Gradient Slingshots' and Section 3.1 'Theoretical Basis', including equations like (2) for the update rule. However, it does not contain any explicitly labeled 'Pseudocode' or 'Algorithm' blocks or figures within the text. |
| Open Source Code | Yes | The Python implementation of GS can be found at: https://github.com/dilyabareeva/grad-slingshot. |
| Open Datasets | Yes | We evaluate pixel-domain FV manipulation [19], referred to as Pixel-AM, using 6-layer CNNs trained on MNIST [58], as interpretability for this FV variant is feasible only for small models. We also assess Fourier FV [35]. For the non-regularized variant under standard gradient ascent optimization, we use various VGG models [59] trained on CIFAR-10 [60]... Res Net-18 [62] trained on Tiny Image Net [63], and Res Net-50 and Vi T-L/32 [64], both pretrained on Image Net-1k [65]. |
| Dataset Splits | Yes | In Table 7, we provide train test splits and preprocessing details for all experimental settings except CLIP Vi T-L/14. For MNIST [58]: 80% / 20%. For CIFAR-10 [60]: Default (80% / 20%). For Tiny Image Net [63]: 80% / 20% of train set. For Image Net [65]: Default; val. used as test. For the toy experiment in Appendix C.1, it states: 'The dataset was partitioned into training and testing subsets, with 128 and 896 data points, respectively.' |
| Hardware Specification | Yes | All experiments were conducted using a workstation with 2 24 GB NVIDIA RTX 4090 GPUs and a compute cluster with 4 40 GB NVIDIA A100 GPUs. Each experiment was performed on a single GPU. |
| Software Dependencies | No | The paper mentions using "lucid and torch-dreams libraries" (Appendix C.6), "torchvision library" (Appendix C.7), and "Adam W [77]" (Appendix C.5). However, it does not specify explicit version numbers for these software components or the programming language (e.g., Python version) and core ML frameworks (e.g., PyTorch version) used, which would be necessary for full reproducibility of software dependencies. |
| Experiment Setup | Yes | Appendix C 'Experimental Details' and its subsections provide specific experimental setup details. For instance, Appendix C.5 specifies optimizer details: 'trained with the SGD optimizer using learning rate of 0.001 and momentum of 0.9 until convergence.' Appendix C.6 in Table 11 and 12 lists Gradient Slingshots hyperparameters such as 'α, w, γ, C, σB, σL' and 'LR, Weight Decay, ϵADAM, Batch Size'. Appendix C.7 describes FV procedure parameters such as 'Step Size, Steps' and details of 'Transformation robustness' including rotation degrees and scaling factors. |