MetaBackdoor shows that LLMs can be backdoored using positional triggers like sequence length, enabling stealthy activation on clean inputs to leak system prompts or trigger malicious behavior.
The Rapid Adoption of 3d Point Cloud Data in Fields Such as Autonomous Vehicles, Robotics, and Virtual Reality Has Spurred Remarkable Improvements in Object Recognition and Scene Understanding. However, These Advancements Also Introduce Serious Security Challenges, Most Notably in the Form of Backdoor Attacks, Which Embed Malicious Triggers into the Training Data to Undermine a Model's Predictions. In this Paper, We Propose Cloudfort, a Novel Defense Mechanism that Strengthens 3d Point Cloud Classifiers Against Backdoor Attacks. Cloudfort Integrates Spatial Partitioning with Ensemble Prediction to Effectively Neutralize Backdoor Triggers While Preserving Classification Accuracy on Benign Samples. We Validate the Efficacy of Cloudfort Through Comprehensive Experiments, Demonstrating its Robust Defense Against the Point Cloud Backdoor Attack. Our Findings Show that Cloudfort Markedly
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CR 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
MetaBackdoor: Exploiting Positional Encoding as a Backdoor Attack Surface in LLMs
MetaBackdoor shows that LLMs can be backdoored using positional triggers like sequence length, enabling stealthy activation on clean inputs to leak system prompts or trigger malicious behavior.