SHIELD: Robust Continual Learning via Hypernetworks and Interval MixUp
The newly introduced framework, SHIELD (Secure Hypernetworks for Incremental Expansion Learning Defense), tackles the difficulties of continual learning in adversarial settings. Detailed in arXiv:2506.08255, SHIELD combines Interval Bound Propagation (IBP) with a hypernetwork architecture to derive task-specific model parameters from compact task embeddings, thus removing the need for replay buffers and complete model replicas. A unique training approach, called Interval MixUp, creates virtual examples as ℓ∞ balls surrounding MixUp points, employing interval arithmetic to ensure certified robustness and smoother decision boundaries. This method has been tested against significant adversarial attacks, demonstrating enhanced robustness and scalability.
Key facts
- SHIELD integrates Interval Bound Propagation with hypernetworks
- Generates task-specific parameters via shared hypernetwork conditioned on task embeddings
- Eliminates need for replay buffers or full model copies
- Introduces Interval MixUp training strategy
- Interval MixUp blends virtual examples as ℓ∞ balls around MixUp points
- Uses interval arithmetic to guarantee certified robustness
- Mitigates the wrapping effect for smoother decision boundaries
- Evaluated under strong adversarial conditions
Entities
Institutions
- arXiv