Devling into Adversarial Transferability on Image Classification: Review, Benchmark, and Evaluation

This document explores adversarial transferability in image classification, a property where adversarial examples generated on one surrogate model can deceive alternate, unexposed victim models without direct access. This phenomenon raises significant security concerns in practical applications and has attracted substantial research attention. The authors identify a critical gap in the field: the absence of a standardized framework and criteria for evaluating transfer-based attacks, which often leads to biased assessments. To address this, they conduct an exhaustive review, categorizing existing transfer-based attacks into six distinct methodologies. A comprehensive framework is then proposed to serve as a benchmark for fairly evaluating these diverse attacks. The paper further delineates common strategies that enhance adversarial transferability and highlights prevalent issues that could lead to unfair comparisons in existing studies. It also provides the most extensive overview to date of over one hundred untargeted and targeted transfer-based attacks. The work aims to inspire more robust attack strategies and promote comprehensive, equitable evaluation criteria in the domain, concluding with a brief review of transfer-based attacks beyond image classification.
#AdversarialAttacks #Transferability #ImageClassification #DeepLearning #Cybersecurity #MachineLearning #Benchmark #ArtificialIntelligence #AI

Donats: https://www.patreon.com/c/luxak

paper - https://arxiv.org/pdf/2602.23117v1
subscribe - https://t.me/arxivpaper
created with NotebookLM

Видео Devling into Adversarial Transferability on Image Classification: Review, Benchmark, and Evaluation канала LuxaK