Target Selection for Adversarial Evaluation of Vision Models Using Language Models - Codes and Results

The repository provides reproducible code and configuration scripts for generating and evaluating similarity‑based adversarial attacks on vision models, using pretrained language and vision‑language models as semantic oracles. It implements end‑to‑end pipelines that select target classes according to their semantic proximity or distance in the joint embedding space, and then craft targeted adversarial examples that preserve human‑perceived meaning while inducing controlled misclassifications. The codebase includes attack routines, model wrappers, and utilities for data preprocessing, logging, and result aggregation, enabling systematic experiments across multiple similarity sources, attack methods, and vision architectures under standardized evaluation settings. The repository is released as an archival research artifact and supplementary material to: Reproduce the experiments reported in "Target Selection for Adversarial Evaluation of Vision Models Using Language Models," Accepted for Presentation at the IEEE WCCI/IJCNN 2026 (21-26 June 2026); Inspect the implementation details of the proposed semantics‑guided target selection framework; and Adapt the pipeline to new datasets or models while preserving comparable protocols. It contains example development files (e.g., ImageNet-1k CSVs and similarity matrices derived from BERT, CLIP, and LLaMA, as well as a WordNet/WuP reference), the main adversarial testing script, configuration files, and result folders, together with instructions to run all experiments in Python 3.9.16 using the provided requirements and the CleverHans library for attack generation. Please read the Readme.md file for further details.