核心概念
Dual Randomized Smoothing (DRS) provides a tight ℓ2 certified robustness radius for high-dimensional inputs by employing dual smoothing in the lower-dimensional space, effectively mitigating the curse of dimensionality.
摘要
This paper introduces a novel smoothing mechanism called Dual Randomized Smoothing (DRS) to provide certified robustness for high-dimensional inputs.
Key highlights:
DRS partitions the original d-dimensional input into two sub-inputs with lower dimensionality of m and n, and performs dual smoothing within the lower-dimensional space.
Theoretically, the paper proves that DRS guarantees a tight ℓ2 certified robustness radius for the original input and reveals that DRS attains a superior upper bound on the ℓ2 robustness radius, which decreases proportionally at a rate of (1/√m + 1/√n) with m + n = d.
Extensive experiments demonstrate the generalizability and effectiveness of DRS. DRS can adeptly integrate with various existing methods, resulting in substantial enhancements to both the accuracy and the certified robustness baseline of Randomized Smoothing (RS).
Compared to RS, DRS consistently increases the classification accuracy significantly while simultaneously improving or preserving the Average Certified Radius (ACR) under lower noise levels. When increasing the noise level, DRS significantly improves the ACR but experiences an accuracy drop, due to the high level of noise compromising the utility of the information.
Applying model ensemble techniques can further boost the performance of DRS, enhancing the certified accuracy and average certified robustness.
統計資料
The upper bound of ℓ2 certified radius (calculated by Equation 4 and Equation 15) of RS and DRS with σ = 1/√d and smoothed probability = 0.999.