New 'Season' framework boosts adversarial attack transferability across AI models

Researchers have developed a new framework called Season to improve the effectiveness of adversarial attacks on image recognition models. This framework specifically addresses the challenge of transferability, where attacks successful on one model architecture often fail on another due to differing feature extraction methods (CNNs focusing on local textures vs. Vision Transformers on global shapes). Season refines gradient updates by separating them into low-frequency (structural) and high-frequency (textural) components, then strategically reallocating energy and ensuring orthogonality to minimize interference. When applied as a plug-and-play wrapper to existing attacks, Season demonstrated an average improvement of 6.6 percentage points in transfer success rates across various CNN, ViT, and MLP architectures. AI

IMPACT Enhances understanding of model vulnerabilities and potential defenses against adversarial attacks across diverse AI architectures.

RANK_REASON The cluster contains a research paper detailing a new technical framework for adversarial attacks. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• CNN
• ImageNet
• multilayer perceptron
• Vision Transformers

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →