New SEP-Attack method enhances transferable adversarial text attacks

Researchers have introduced SEP-Attack, a novel method for generating adversarial text attacks that are transferable to different models. This approach utilizes Determinantal Point Process to create diverse ensemble weights, improving the representation of submodel transferability. SEP-Attack also employs a new metric for evaluating prediction confidence to better estimate word importance and generate adversarial candidates, outperforming existing methods on multiple datasets and real-world APIs. AI

IMPACT This research introduces a more effective method for generating transferable adversarial text attacks, potentially improving the robustness and security of NLP models.

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

Read on arXiv cs.AI →

• Determinantal Point Process
• SEP-Attack

• paper
• safety

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