New framework RPP detects poisoned samples in backdoor attacks

Researchers have developed a new framework called Randomized Probability Perturbation (RPP) to detect poisoned samples used in backdoor attacks against deep neural networks. This framework is designed to be effective even when datasets are imbalanced, a common issue in real-world scenarios that can exacerbate backdoor vulnerabilities. RPP operates in a black-box setting, relying solely on model output probabilities to identify manipulated inputs and offering provable guarantees on detectability. AI

IMPACT This research could lead to more robust defenses against sophisticated attacks that exploit dataset imbalances in real-world AI applications.

RANK_REASON This is a research paper detailing a new framework for detecting security vulnerabilities in machine learning models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• CIFAR-10
• ImageNet10
• Miao Lin
• MNIST database
• Randomized Probability Perturbation
• The Street View House Numbers Dataset
• TinyImageNet

• paper
• safety

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