TCAP: Tri-Component Attention Profiling for Unsupervised Backdoor Detection in MLLM Fine-Tuning
📰 ArXiv cs.AI
Learn to detect backdoors in fine-tuned MLLMs using TCAP, a novel unsupervised method that profiles attention allocation divergence
Action Steps
- Implement TCAP to profile attention allocation in fine-tuned MLLMs
- Analyze attention allocation divergence to detect potential backdoors
- Use the detected backdoors to retrain the model with cleaned data
- Evaluate the effectiveness of TCAP in detecting backdoors across diverse trigger types and modalities
- Integrate TCAP into the fine-tuning pipeline to prevent backdoor attacks
Who Needs to Know This
ML engineers and researchers working with MLLMs can benefit from this technique to ensure the security of their fine-tuned models
Key Insight
💡 Attention allocation divergence can be used as a universal backdoor fingerprint for unsupervised detection
Share This
🚨 Detect backdoors in fine-tuned MLLMs with TCAP! 🚨
Key Takeaways
Learn to detect backdoors in fine-tuned MLLMs using TCAP, a novel unsupervised method that profiles attention allocation divergence
Full Article
Title: TCAP: Tri-Component Attention Profiling for Unsupervised Backdoor Detection in MLLM Fine-Tuning
Abstract:
arXiv:2601.21692v2 Announce Type: replace Abstract: Fine-Tuning-as-a-Service (FTaaS) facilitates the customization of Multimodal Large Language Models (MLLMs) but introduces critical backdoor risks via poisoned data. Existing defenses either rely on supervised signals or fail to generalize across diverse trigger types and modalities. In this work, we uncover a universal backdoor fingerprint-attention allocation divergence-where poisoned samples disrupt the balanced attention distribution across
Abstract:
arXiv:2601.21692v2 Announce Type: replace Abstract: Fine-Tuning-as-a-Service (FTaaS) facilitates the customization of Multimodal Large Language Models (MLLMs) but introduces critical backdoor risks via poisoned data. Existing defenses either rely on supervised signals or fail to generalize across diverse trigger types and modalities. In this work, we uncover a universal backdoor fingerprint-attention allocation divergence-where poisoned samples disrupt the balanced attention distribution across
DeepCamp AI