
The author proposes using model merging as an effective defense against backdoor attacks on language models, showcasing robustness and versatility in various contexts.


coremsg

Mitigating-Backdoor-Attacks-on-Language-Models-through-Model-Merging


Mitigating Backdoor Attacks on Language Models through Model Merging


title_rewrite


Venom proposes a novel approach to enhance the survivability of backdoor attacks against model reconstruction-based defenses by coupling decision paths, preserving attack capabilities while improving survivability.


Enhancing-Backdoor-Attack-Survivability-with-Venom


Enhancing Backdoor Attack Survivability with Venom



A lightweight defense mechanism, PAD-FT, that effectively disinfects poisoned deep neural network models without requiring additional clean data.


pad-ft-a-lightweight-defense-against-backdoor-attacks-via-data-purification-and-fine-tuning


PAD-FT: A Lightweight Defense Against Backdoor Attacks via Data Purification and Fine-Tuning
