New data attack method preserves image quality while poisoning deep learning models

Researchers have developed a new data availability attack (DAA) called Imperfect Restoration Poisoning (IRP) that aims to make data unlearnable for deep learning models. Existing DAAs struggle with a trade-off between image quality and poisoning effectiveness, particularly against self-supervised learning (SSL) methods. IRP addresses these limitations by preserving high image quality while achieving strong poisoning effects, outperforming eight baseline attacks and five defense methods in extensive comparisons. AI

IMPACT This research introduces a more effective method for poisoning training data, potentially impacting the integrity and reliability of deep learning models.

RANK_REASON Research paper detailing a new method for data poisoning attacks on deep learning models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• Convolution-based Unlearnable Dataset
• CUDA
• Hugging Face
• Imperfect Restoration Poisoning
• Self-Supervised Learning
• Supervised Learning

• paper
• safety

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