Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

BadPromptFL: A Novel Backdoor Threat to Prompt-based Federated Learning in Multimodal Models

Created by
  • Haebom

Author

Maozhen Zhang, Mengnan Zhao, Bo Wang

Outline

This paper presents BadPromptFL, a novel backdoor attack on prompt-based federated learning (PromptFL) in multimodal contrastive learning models. BadPromptFL injects malicious prompts into the global aggregation process by having compromised clients jointly optimize local backdoor triggers and prompt embeddings. These malicious prompts are then propagated to benign clients, enabling universal backdoor activation during inference without modifying model parameters. Leveraging the contextual learning behavior of a CLIP-style architecture, BadPromptFL achieves a high attack success rate (e.g., >90%) with minimal visibility and limited client involvement. Extensive experiments on diverse datasets and aggregation protocols demonstrate the attack's effectiveness, stealth, and generalizability, raising serious concerns about the robustness of prompt-based federated learning in real-world deployments.

Takeaways, Limitations

•
Takeaways: We first uncovered a security vulnerability in prompt-based federated learning and presented an effective and stealthy backdoor attack technique called BadPromptFL, highlighting the need to strengthen the security of prompt-based federated learning systems in real-world environments. The attack technique, which exploits the characteristics of the CLIP-style architecture, suggests its applicability to other similar models.
•
Limitations: The defense techniques against the currently proposed BadPromptFL attack are not covered in this paper. Further research on various defense techniques is needed. Since these results are experimental results for a specific dataset and model architecture, further research is needed to determine their generalizability to other environments.
👍
Made with Slashpage