PulseAugur
EN
LIVE 10:59:50

New DFCS method boosts backdoor attack efficiency by 4.60% · 2 sources tracked

By PulseAugur Editorial · [2 sources] ·

Researchers have developed a new method called Distributional Feature Coverage Sample Selection (DFCS) to improve the efficiency of backdoor attacks on machine learning models. This training-free, trigger-agnostic approach clusters pre-trained features and selects the nearest sample from each cluster, aiming to avoid redundant data points. DFCS demonstrated a high average attack success rate of 96.30% across various datasets and attack types, outperforming existing methods by an average of 4.60 percentage points while maintaining clean accuracy. AI

IMPACT Enhances the effectiveness of data-efficient backdoor attacks, potentially impacting model security and data integrity.

RANK_REASON The cluster describes a new method proposed in a research paper for improving backdoor attacks on machine learning models.

Read on arXiv cs.CV →

AI-generated summary · Google Gemini · from 2 sources. How we write summaries →

New DFCS method boosts backdoor attack efficiency by 4.60% · 2 sources tracked

COVERAGE [2]

  1. Hugging Face Daily Papers TIER_1 English(EN) ·

    Diversity Matters: Distributional Feature Coverage Sample Selection for Data-Efficient Backdoor Attacks

    Backdoor attacks compromise training data so that a model retains clean accuracy but predicts an attacker-chosen target on triggered inputs. At very low poisoning rates, only a few samples convey the trigger--target association, making poison-sample selection critical. Existing m…

  2. arXiv cs.CV TIER_1 English(EN) · Yi Yang, Xiaoke Chen, Jinyang Huang, Feng-Qi Cui, Yu-Tong Guo, Jia-Cheng Zhao, Haiming Jin, Xiaokang Zhou, Meng Li ·

    Diversity Matters: Distributional Feature Coverage Sample Selection for Data-Efficient Backdoor Attacks

    arXiv:2608.09047v1 Announce Type: cross Abstract: Backdoor attacks compromise training data so that a model retains clean accuracy but predicts an attacker-chosen target on triggered inputs. At very low poisoning rates, only a few samples convey the trigger--target association, m…

RELATED ENTITIES

RELATED TOPICS