Faithful Knowledge Distillation
This addresses the need for reliable deployment of compressed networks in safety-critical settings, though it is incremental as it builds on existing adversarial robustness work in knowledge distillation.
The paper tackles the problem of ensuring that a student network in knowledge distillation maintains similar confidence levels and agreement with its teacher near correctly classified examples, particularly for safety-critical applications, by introducing a faithful distillation framework that improves verifiability on datasets like MNIST, Fashion-MNIST, and CIFAR-10.
Knowledge distillation (KD) has received much attention due to its success in compressing networks to allow for their deployment in resource-constrained systems. While the problem of adversarial robustness has been studied before in the KD setting, previous works overlook what we term the relative calibration of the student network with respect to its teacher in terms of soft confidences. In particular, we focus on two crucial questions with regard to a teacher-student pair: (i) do the teacher and student disagree at points close to correctly classified dataset examples, and (ii) is the distilled student as confident as the teacher around dataset examples? These are critical questions when considering the deployment of a smaller student network trained from a robust teacher within a safety-critical setting. To address these questions, we introduce a faithful imitation framework to discuss the relative calibration of confidences and provide empirical and certified methods to evaluate the relative calibration of a student w.r.t. its teacher. Further, to verifiably align the relative calibration incentives of the student to those of its teacher, we introduce faithful distillation. Our experiments on the MNIST, Fashion-MNIST and CIFAR-10 datasets demonstrate the need for such an analysis and the advantages of the increased verifiability of faithful distillation over alternative adversarial distillation methods.