المفاهيم الأساسية
This paper proposes a novel framework called RSCP+ to provide provable robustness guarantee for conformal prediction methods against adversarial attacks. It also introduces two new methods, Post-Training Transformation (PTT) and Robust Conformal Training (RCT), to effectively reduce the size of prediction sets with little computational overhead.
الملخص
The paper addresses two key limitations of the previous Randomized Smoothed Conformal Prediction (RSCP) method:
Robustness Guarantee: The authors identify a flaw in the robustness certification of RSCP and propose a new scheme called RSCP+ to provide provable robustness guarantee in practice.
RSCP uses randomized smoothing to construct a new non-conformity score that is robust to adversarial perturbations. However, RSCP's robustness guarantee is flawed when Monte Carlo sampling is used for randomized smoothing, which is the common practice.
To address this, the authors propose RSCP+ which directly uses the Monte Carlo estimator as the base score and derives a new robustness guarantee.
Efficiency: The authors show that directly applying RSCP+ often leads to trivial prediction sets that give the entire label set, due to the conservativeness of RSCP.
To improve efficiency, the authors propose two new methods:
Post-Training Transformation (PTT): A scalable, training-free method that applies a two-step transformation (ranking and sigmoid) on the base score to reduce the conservativeness.
Robust Conformal Training (RCT): A general training framework that incorporates the RSCP+ process into the training of the base classifier to further boost efficiency.
The experimental results on CIFAR10, CIFAR100 and ImageNet demonstrate that the baseline method only yields trivial predictions, while the authors' proposed methods can boost the efficiency by up to 4.36×, 5.46×, and 16.9× respectively, while providing practical robustness guarantee.
الإحصائيات
The average size of prediction sets on CIFAR10 is reduced from 10 (baseline) to 2.294 (PTT) and 2.294 (PTT+RCT).
The average size of prediction sets on CIFAR100 is reduced from 100 (baseline) to 26.06 (PTT) and 18.30 (PTT+RCT).
The average size of prediction sets on ImageNet is reduced from 1000 (baseline) to 94.66 (PTT) and 59.12 (PTT+Bernstein).