Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
AI Research
arXiv CS.AI·24d ago·5 min read
Perturbing the Phase: Analyzing Adversarial Robustness of Complex-Valued Neural Networks
TL;DR
Complex-valued neural networks (CVNNs) exhibit unique vulnerabilities to adversarial attacks, particularly through phase perturbations. Phase Attacks, targeting phase information, reveal that CVNNs can be more robust than real-valued networks (RVNNs) in some cases, but are still highly susceptible to performance degradation from phase changes.
✦ Why It Matters
Consider implementing Phase Attacks in your testing to better understand CVNN vulnerabilities.
Key Takeaways
- Explore the robustness of CVNNs against phase perturbations.
- Implement Phase Attacks to evaluate model vulnerabilities.
- Adapt existing adversarial techniques for complex-valued inputs.
- Consider phase information in the design of CVNNs.
How It Works
Phase Attacks specifically target the phase component of complex-valued inputs, which is crucial for the performance of CVNNs. By manipulating this phase information, the attacks can lead to significant drops in model accuracy, revealing vulnerabilities that are not as pronounced in traditional real-valued networks.