Evaluating NEO-KD Against Single-Exit Defense Methods in Multi-Exit Networks

cover
30 Sept 2024

Authors:

(1) Seokil Ham, KAIST;

(2) Jungwuk Park, KAIST;

(3) Dong-Jun Han, Purdue University;

(4) Jaekyun Moon, KAIST.

Abstract and 1. Introduction

2. Related Works

3. Proposed NEO-KD Algorithm and 3.1 Problem Setup: Adversarial Training in Multi-Exit Networks

3.2 Algorithm Description

4. Experiments and 4.1 Experimental Setup

4.2. Main Experimental Results

4.3. Ablation Studies and Discussions

5. Conclusion, Acknowledgement and References

A. Experiment Details

B. Clean Test Accuracy and C. Adversarial Training via Average Attack

D. Hyperparameter Tuning

E. Discussions on Performance Degradation at Later Exits

F. Comparison with Recent Defense Methods for Single-Exit Networks

G. Comparison with SKD and ARD and H. Implementations of Stronger Attacker Algorithms

F Comparison with Recent Defense Methods for Single-Exit Networks

The baselines in the main paper were generally the adversarial defense methods designed for multi-exit networks. In this section, we conduct additional experiments with a recent defense method, TEAT [7], and compare with our method. Since TEAT was originally designed for the single-exit network, we first adapt TEAT to the multi-exit network setting. Instead of the original TEAT that generates the adversarial examples considering the final output of the network, we modify TEAT to generate adversarial examples that maximizes the average loss of all exits in the multi-exit network. Table A3 below shows the results using max-average attack on CIFAR-10/100. It can be seen that our NEO-KD, which is designed for multi-exit networks, achieves higher adversarial test accuracy compared to the TEAT methods (PGD-TE and TRADES-TE) designed for single-exit networks. The results highlight the necessity of developing adversarial defense techniques geared to multi-exit networks rather than adapting general defense methods used for single-exit networks.

Table A3: Comparison of adversarial test accuracy against max-average attack between TEAT methods and our NEO-KD.

Table A4: Adversarial test accuracy of SKD and ARD according to exit selection as a teacher prediction.

This paper is available on arxiv under CC 4.0 license.