HyCAS is a randomized Lipschitz-constrained convolutional architecture that improves both certified L2 accuracy and empirical robustness against strong attacks on image benchmarks without losing clean accuracy.
Independently of input noise, we certify the backbone + internal randomness by averaging logits only over Ω: Z(x) ≜ EΩ [ s(x; Ω) ], with Lip(Z) ≤ 2
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Certified vs. Empirical Adversarial Robust-ness via Hybrid Convolutions with Attention Stochasticity
HyCAS is a randomized Lipschitz-constrained convolutional architecture that improves both certified L2 accuracy and empirical robustness against strong attacks on image benchmarks without losing clean accuracy.