Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
AI Research
arXiv CS.AI·24d ago·5 min read
Information-Geometric Inverse Distillation for Enhancing Adversarial Transferability
TL;DR
Inverse Knowledge Distillation (IKD) enhances adversarial transferability by maximizing the prediction discrepancy between benign and adversarial samples in surrogate models. Extensive experiments show consistent performance improvements across various architectures, establishing IKD as a lightweight and effective method.
✦ Why It Matters
Implement IKD in your adversarial attack strategies to enhance transferability and effectiveness.
Key Takeaways
- Integrate IKD into existing adversarial attack frameworks.
- Utilize soft-label objectives for better prediction discrepancies.
- Focus on Fisher information for improved model transferability.
- Conduct experiments on diverse architectures for validation.
How It Works
IKD operates by maximizing the prediction distribution discrepancy between benign and adversarial samples. It uses a soft-label objective that aligns with cross-entropy and Kullback-Leibler divergence, effectively pushing adversarial predictions away from a fixed benign prediction anchor.
This approach enriches the attack with Fisher-sensitive directions, which are crucial for enhancing transferability across different models.