Understanding Diffusion Models Towards Adversarial Robustness

5 minute read

1. Introduction

  • Diffusion model (์ดํ•˜ DM)์€ likelihood-based ๋ชจ๋ธ๋กœ, ์ƒ์„ฑํ•˜๋Š” ์ƒ˜ํ”Œ ํ€„๋ฆฌํ‹ฐ๊ฐ€ ์ข‹์€ ๋™์‹œ์— GAN์ด ๊ฐ€์ง€๊ณ  ์žˆ๋˜ mode collapse issue๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ํ•ด๊ฒฐํ•˜๊ณ  ์žˆ๋‹ค.

  • DM์€ input image์— ๋…ธ์ด์ฆˆ๋ฅผ ์ฃผ๊ธฐ์ ์œผ๋กœ ์ถ”๊ฐ€ํ•˜๋Š” forward process์™€ ๋…ธ์ด์ฆˆ๋ฅผ ์ œ๊ฑฐํ•ด ๋‚˜๊ฐ€๋Š” reverse process๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ๋‹ค.

  • Empirical research๋ฅผ ํ†ตํ•ด DM์ด adversarial attack์„ ํšจ๊ณผ์ ์œผ๋กœ ์ œ๊ฑฐํ•œ๋‹ค๋Š” ๊ฒƒ์ด ๋‚˜ํƒ€๋‚ฌ์ง€๋งŒ ์ด์— ๋Œ€ํ•œ ๊ตฌ์ฒด์ ์ธ ์ด์œ ์— ๋Œ€ํ•ด์„œ๋Š” ์•„์ง ์—ฐ๊ตฌ๋˜์ง€ ์•Š์•˜๋‹ค.

  • ์ด ๋…ผ๋ฌธ์—์„œ๋Š” DM์ด adversarial attack์„ 1. ์–ด๋–ป๊ฒŒ ํšจ๊ณผ์ ์œผ๋กœ ์ œ๊ฑฐํ•˜๋Š”์ง€์™€ ์ด๋ฅผ ํ†ตํ•ด 2. DM์„ ํ†ตํ•ด ๋ชจ๋ธ์„ robustํ•˜๊ฒŒ ๋งŒ๋“œ๋Š” ์ƒˆ๋กœ์šด framework์ธ DensePure์„ ์ œ์‹œํ•œ๋‹ค.

2. Preliminaries and Backgrounds

1. Continuous-Time Diffusion Model

โ€˜Score-based generative modeling through stochastic differential equations.โ€™ ๋…ผ๋ฌธ์—์„œ๋Š” Score-based generative model๋“ค์€ SDE(Stochastic Differential Equation) framework์•ˆ์—์„œ ํฌ๊ฒŒ ๋‘๊ฐ€์ง€ ์š”์†Œ์ธ diffusion forward process์™€ reverse process๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ๋‹ค๊ณ  ์ฃผ์žฅํ•˜๊ณ  ์žˆ๋‹ค. ๋จผ์ € forward diffusion process์€ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ๋‚˜ํƒ€๋‚ผ ์ˆ˜ ์žˆ๋‹ค.

๋‹ค์Œ์œผ๋กœ reverse process๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์ด reverse-time SDE๋กœ ๋‚˜ํƒ€๋‚ผ ์ˆ˜ ์žˆ๋‹ค.


2. Discrete-Time Diffusion Model(or DDPM)

โ€˜Denoising Diffusion Probablistic Modelโ€™๋…ผ๋ฌธ์—์„œ๋Š” DDPM์˜ forward diffusion process๋ฅผ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ๋‚˜ํƒ€๋‚ธ๋‹ค.

Reverse diffusion process๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์ด ๋‚˜ํƒ€๋‚ธ๋‹ค.

์—ฌ๊ธฐ์„œ ์ด ๋…ผ๋ฌธ์—์„œ๋Š” ๋‹ค์Œ step์—์„œ ๋…ธ์ด์ฆˆ๋ฅผ ์ œ๊ฑฐํ•œ ์ƒ˜ํ”Œ ๋ถ„ํฌ์˜ ๋ถ„์‚ฐ์„ $\beta_i$ ($i$๋ฒˆ์งธ step์— ์ฃผ์ž…ํ•œ ๋…ธ์ด์ฆˆ์˜ ์–‘) ์œผ๋กœ ๊ณ ์ •ํ•˜๊ณ , ํ‰๊ท ์ธ $\mu_\theta (x_i, i)$ ๋งŒ์„ ๊ตฌํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๊ฐ„๋‹จํžˆ ํ•˜์˜€๋‹ค. ์†์‹ค ํ•จ์ˆ˜๋ฅผ ๊ณ„์‚ฐํ•˜๋Š” ๊ณผ์ •์ด ์›Œ๋‚™ ๋ณต์žกํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ด ๊ณผ์ •์„ ์ƒ๋žตํ•˜๊ณ  ์ตœ์ข…์ ์ธ ์†์‹ค ํ•จ์ˆ˜๋งŒ์„ ๋‚˜ํƒ€๋‚ด๋ฉด ๋‹ค์Œ๊ณผ ๊ฐ™์ด ๋‚˜ํƒ€๋‚ผ ์ˆ˜ ์žˆ๋‹ค.

์œ„ ๊ณต์‹์€ ๊ณง $i$๋ฒˆ์งธ step์—์„œ ๋…ธ์ด์ฆˆ๋ฅผ ์ถ”๊ฐ€ํ•œ ์ด๋ฏธ์ง€๋ฅผ ์ž…๋ ฅ ๊ฐ’์œผ๋กœ ๋„ฃ์—ˆ์„ ๋•Œ ์ถ”๊ฐ€ํ•œ ๋…ธ์ด์ฆˆ์˜ ์–‘์„ ์ถœ๋ ฅํ•˜๋Š” ๋„คํŠธ์›Œํฌ๋ฅผ ๊ตฌ์„ฑํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•œ๋‹ค๋Š” ๊ฒƒ์œผ๋กœ ์ดํ•ดํ•  ์ˆ˜ ์žˆ๋‹ค.


3. Randomized Smoothing

RS(Randomized Smoothing)์€ $L_2$-norm์— ํ•ด๋‹น๋˜๋Š” adversarial attack์„ ํšจ๊ณผ์ ์œผ๋กœ ๋ง‰๊ธฐ ์œ„ํ•ด ๊ณ ์•ˆ๋œ ๋ฐฉ๋ฒ•์ด๋‹ค. ๋‹ค์Œ๊ณผ ๊ฐ™์€ ๋ฐฉ๋ฒ•์œผ๋กœ classifier์„ smoothing ํ•ด์คŒ์œผ๋กœ์จ classifier์„ robustํ•˜๊ฒŒ ๋งŒ๋“ค์–ด์ค€๋‹ค.

์—ฌ๊ธฐ์„œ $\sigma$๋Š” ๋ชจ๋ธ์˜ robustness์™€ accuracy๋ฅผ ์กฐ์ ˆํ•˜๋Š” parameter์ด๋‹ค. ์•„์ง ํ•ด๋‹น ๋…ผ๋ฌธ์„ ์ฝ์–ด๋ณด์ง€ ์•Š์•˜๊ธฐ ๋•Œ๋ฌธ์— ์ž์„ธํ•œ ์ดํ•ด๋ฅผ ์œ„ํ•ด ์ถ”๊ฐ€๋กœ ์ฝ์–ด ๋ณผ ์˜ˆ์ •์ด๋‹ค.

3. Theoretical Analysis

ํ•ด๋‹น ์„น์…˜์—์„œ๋Š” DM์ด ์–ด๋–ป๊ฒŒ adversarial attack์„ ํšจ๊ณผ์ ์œผ๋กœ ์ œ๊ฑฐํ•˜๋Š”์ง€์— ๋Œ€ํ•ด ๋‹ค์Œ๊ณผ ๊ฐ™์€ 3๊ฐ€์ง€ Theorem์„ ํ†ตํ•ด ๋ฐํ˜€๋‚ด๊ณ  ์žˆ๋‹ค. ๋จผ์ € ํ•ด๋‹น ์„น์…˜์˜ Theorem๋“ค์€ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ๊ฐ€์ •์„ ๋งŒ์กฑํ•˜๊ณ  ์žˆ๋‹ค.

Theorem 3.1

Perturbed๋œ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ์ธ $x_{a,t}$๊ฐ€ reverse-SDE๋ฅผ ๊ฑฐ์น˜๊ณ  ๋‚œ ํ›„ purify๋œ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ $\hat{x}$ ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์กฐ๊ฑด๋ถ€ ํ™•๋ฅ ๋ถ„ํฌ๋ฅผ ๊ฐ€์ง„๋‹ค.

์œ„ ๊ณต์‹์„ ํ†ตํ•ด $\mathbb{P}$๋ฅผ ์ฆ๊ฐ€์‹œํ‚ค๊ธฐ ์œ„ํ•ด์„œ๋Š” $\left | x-x_a \right |_2^2$๋ฅผ ๊ฐ์†Œ์‹œ์ผœ์•ผ ํ•˜๊ณ , ์ด๋Š” ๊ณง $x$๊ฐ€ $x_a$ ์ฃผ๋ณ€์œผ๋กœ ๋†’์€ ๋ฐ€๋„๋ฅผ ๊ฐ€์ ธ์•ผ ํ•จ์„ ์•Œ ์ˆ˜ ์žˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋‹ค์Œ๊ณผ ๊ฐ™์ด $\mathbb{P}$๋ฅผ ์ตœ๋Œ€ํ™”ํ•˜๋Š” ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ $x$๋ฅผ ์„ ํƒํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๊ฐ„๋‹จํžˆ ๋‚˜ํƒ€๋‚ผ ์ˆ˜ ์žˆ๋‹ค.

Theorem 3.2

$f$๊ฐ€ classifier์ด๊ณ  $G(x_0)$๊ฐ€ $x_0$๊ณผ ๊ฐ™์€ class๋ฅผ ๊ฐ€์ง€๋Š” data region์ด๋ผ๊ณ  ํ•˜์ž. ์—ฌ๊ธฐ์„œ $P(\cdot ;\psi)$๊ฐ€ purification model ์ด๋ผ๊ณ  ํ•  ๋•Œ $G(x_0)$์˜ robust data region์€ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์ •์˜ํ•  ์ˆ˜ ์žˆ๋‹ค.

์œ„ ์‹์€ purified ๋œ $x$๊ฐ€ $x_0$๊ณผ ๊ฐ™์€ label์„ ๊ฐ€์ง€๋Š” $x$์˜ ์ง‘ํ•ฉ์„ ๋‚˜ํƒ€๋‚ด๋Š” robust data region์ด๋ผ๊ณ  ํ•ด์„ํ•  ์ˆ˜ ์žˆ๋‹ค.
๋˜ํ•œ $x_0$์˜ robust radius๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์ •์˜ํ•  ์ˆ˜ ์žˆ๋‹ค.

์œ„ ์‹์€ ๊ณง $x_0$์„ ๋‘˜๋Ÿฌ์‹ธ๊ณ  ์žˆ๋Š” $D(x_0 ; \psi)$ ์˜ maximum inclined ball์˜ ๋ฐ˜์ง€๋ฆ„์ด๋‹ค.

๊ฒฐ๊ตญ Theorem 3.2๋ฅผ ํ†ตํ•ด ๋‚˜ํƒ€๋‚ด๊ณ  ์‹ถ์€ ๊ฒƒ์€ ๋‹ค์Œ๊ณผ ๊ฐ™๋‹ค:
$x_a$๊ฐ€ Euculidean distance ์ƒ์œผ๋กœ $x_0$๊ณผ ์ถฉ๋ถ„ํžˆ ๊ฐ€๊นŒ์šธ ๋•Œ $x_a$๋Š” $x_0$์™€ purified ๋œ sample์ธ $P(x_a;t)$๊ณผ ๊ฐ™์€ label semantics๋ฅผ ์œ ์ง€ํ•˜๊ณ  ๊ฐ™์€ label๋กœ ๋ถ„๋ฅ˜ ํ•œ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ $x_a$๊ฐ€ $x_0$์™€ ๊ฐ€๊น์ง€ ์•Š์•„๋„ $x_0$๊ณผ ๊ฐ™์€ label์„ ๊ฐ€์ง€๊ณ  ์žˆ๋Š” ๋‹ค๋ฅธ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ, $ \tilde{x}$์™€ ๊ฐ€๊นŒ์›Œ๋„ $x_0$์™€ ๊ฐ™์€ label๋กœ ๋ถ„๋ฅ˜ํ•œ๋‹ค. ์ด๊ฒƒ์— ๋Œ€ํ•œ ์ฆ๋ช…์„ ๋‹ค์Œ Theorem์—์„œ ํ•œ๋‹ค.

Theorem 3.3

1. $x_0$๊ฐ€ ground truth label์„ ๊ฐ€์ง€๊ณ  ์žˆ๋Š” data point์ด๊ณ  $x_a$๊ฐ€ perturbed ๋œ $x_0$๋ผ๋ฉด purified๋œ $P(x_a;t)$๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์€ convex set์— ํฌํ•จ๋  ๋•Œ $x_0$์™€ ๊ฐ™์€ label์„ ๊ฐ€์ง„๋‹ค.

2. ๋˜ํ•œ $x_a$๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์€ convex set์— ํฌํ•จ๋  ๋•Œ $x_0$๊ณผ ๊ฐ™์€ label์„ ๊ฐ€์ง„๋‹ค.

์—ฌ๊ธฐ์„œ 1๊ณผ 2์˜ ๋‹ค๋ฅธ ์ ์€ 2๋Š” $x_0$๊ณผ ๊ฐ™์€ ๋ ˆ์ด๋ธ”์„ ๊ฐ€์ง„ ๋‹ค๋ฅธ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ $\tilde{x}$๋ฅผ ์‚ฌ์šฉํ•œ๋‹ค๋Š” ์ ์ด๋‹ค.

๊ฒฐ๊ตญ robust radius์ธ $r(G(x_0);t$๋ฅผ ์ฐพ๋Š” ๊ฒƒ์ด ์ด ๋ฌธ์ œ์˜ ํ•ต์‹ฌ์ด๋ผ๊ณ  ํ•  ์ˆ˜ ์žˆ๋‹ค. ์—ฌ๊ธฐ์„œ ํ•œ๊ฐ€์ง€ ์ฃผ์˜ํ•ด์•ผ ํ•  ์ ์ด ์žˆ๋‹ค. ๋ฐ”๋กœ $D_{sub}(x_0 ; t)$๋Š” convex์—ฌ๋„ $D(G(x_0);t)$๋Š” convex๊ฐ€ ์•„๋‹ˆ๋ผ๋Š” ์ ์ด๋‹ค. ๋”ฐ๋ผ์„œ $D(G(x_0);t)$๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์€ ๋ฐฉ๋ฒ•์œผ๋กœ ๋ฌธ์ œ์— ์ ‘๊ทผํ•ด์•ผ ํ•œ๋‹ค.

  1. non-convex optimization ๋ฐฉ๋ฒ•์„ ์ด์šฉํ•ด์•ผ ํ•œ๋‹ค.
  2. $D_{sub}(x_0 ; t)$๋Š” convex์ด๋ฏ€๋กœ convex optimization์„ ์ด์šฉํ•ด ํ•ด๊ฒฐํ•ด์„œ $r(G(x_0);t$์— ๋Œ€ํ•œ lower bound๋ฅผ ์ฐพ์„ ์ˆ˜ ์žˆ๋‹ค.

์ด ๋…ผ๋ฌธ์—์„œ๋Š” 2๋ฒˆ์˜ ๋ฐฉ๋ฒ•์œผ๋กœ ์ ‘๊ทผํ•˜๊ณ  ์žˆ๋‹ค. (๊ทธ ์ด์œ ์— ๋Œ€ํ•ด์„œ ๊ฐ„๋‹จํ•˜๊ฒŒ ์„ค๋ช…ํ•˜๊ณ  ์žˆ์œผ๋‚˜ convex optimization์— ๋Œ€ํ•ด์„œ ๋”ฐ๋กœ ๊ณต๋ถ€๋ฅผ ํ•˜์ง€ ์•Š์•„ ์ดํ•ด๋ฅผ ํ•˜์ง€ ๋ชปํ–ˆ๋‹ค. ์ถ”ํ›„์— ๊ณต๋ถ€ํ•œ ํ›„ ์ด ๋ถ€๋ถ„์— ๋Œ€ํ•ด์„œ๋Š” ๋‹ค์‹œ ์ž‘์„ฑํ•˜๋„๋ก ํ•˜๊ฒ ๋‹ค.) ๊ทธ๋Ÿฌ๋‚˜ $D(G(x_0);t)$๋Š” ๋‹ค๋ฅธ sub region๋“ค์„ ํ•ฉ์ง‘ํ•ฉ ํ•œ ๊ฒƒ์ด๋ฏ€๋กœ sub region๋ณด๋‹ค ํ›จ์”ฌ ๋” ์ปค์งˆ ์œ„ํ—˜์ด ์žˆ๋‹ค. ์ด๊ฒƒ์— ๋Œ€ํ•ด์„œ ํ•ด๋‹น ๋…ผ๋ฌธ์—์„œ๋Š” ์•„๋ž˜์™€ ๊ฐ™์€ figure์„ ํ†ตํ•ด ์„ค๋ช…ํ•˜๊ณ  ์žˆ๋‹ค.

๋”ฐ๋ผ์„œ reverse-SDE๋ฅผ ํ†ตํ•ด ์ •ํ™•ํ•œ ๊ฐ’์„ ์ฐพ๋Š” ๊ฒƒ ๋Œ€์‹ ์— approximation ๋ฐฉ๋ฒ•์„ ์ด์šฉํ•˜๊ณ  ์žˆ๋‹ค. approximation ๋ฐฉ๋ฒ•์—๋Š” ์˜ˆ์‹œ๋กœ score-based model์„ ์‚ฌ์šฉํ•ด reverse-SDE์™€์˜ KL-Divergence๋ฅผ ๋‚˜ํƒ€๋‚ธ ๊ฒƒ์ด Theorem 3.4์ด๋‹ค.
์—ฌ๊ธฐ์„œ convex์— ๋Œ€ํ•ด ์•„๋Š” ๊ฒƒ์ด ๊ฑฐ์˜ ์—†์—ˆ๊ธฐ ๋•Œ๋ฌธ์— convexity๋ฅผ ๋ณด์กดํ•˜๋Š” ์—ฐ์‚ฐ์— ๋Œ€ํ•ด์„œ ๋ณ„๋„๋กœ ์ฐพ์•„๋ณด์•˜๋‹ค. convex set์˜ convexity๋ฅผ ๋ณด์กดํ•˜๋Š” ์—ฐ์‚ฐ์€ ๋‹ค์Œ๊ณผ ๊ฐ™๋‹ค.

  1. Intersection
  2. Scaling and transition
    ์˜ˆ์‹œ) $C$๊ฐ€ convex set์ด๊ณ  $a$, $b$๊ฐ€ ๊ฐ๊ฐ scaling, transition scalar factor์ด๋ฉด $aC+b$ ๋˜ํ•œ convex set์ด๋‹ค.
  3. Affine images and preimages
    ์˜ˆ์‹œ) $f(x)=Ax+b$์ด๊ณ  C๊ฐ€ convex set์ด๋ฉด $f(C)$๋„ convex set์ด๋‹ค. ๋˜ํ•œ D๊ฐ€ convex set์ด๋ฉด $f^{-1}(D)$ ๋˜ํ•œ convex set์ด๋‹ค.

Theorem 3.4

$\lVert {\hat{x_\gamma} \rVert }_{\gamma \in [0, t]}$ ์™€$\lVert {x^{\theta}_\gamma \rVert }_{\gamma \in [0, t]}$๊ฐ€ ๊ฐ๊ฐ reverse-SDE, score-based diffusion model์ด๋ผ๊ณ  ํ•˜๋ฉด ์ด ๋‘ ๋ถ„ํฌ์˜ KL-Divergence๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์ด ๋‚˜ํƒ€๋‚ผ ์ˆ˜ ์žˆ๋‹ค.

DensePure

์ง€๊ธˆ๊นŒ์ง€ ์ฆ๋ช…ํ•œ ๋‚ด์šฉ์„ ๋ฐ”ํƒ•์œผ๋กœ DM์„ ํ†ตํ•ด ๋ชจ๋ธ์„ robustํ•˜๊ฒŒ ๋งŒ๋“œ๋Š” ์ƒˆ๋กœ์šด framework์ธ DensePure ์„ ์ƒˆ๋กœ ์ œ์•ˆํ•˜๊ณ  ์žˆ๋‹ค. ์ด framework๊ฐ€ ์ž‘๋™ํ•˜๋Š” ์ˆœ์„œ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™๋‹ค.

  1. ์ž…๋ ฅ๊ฐ’ $x$๊ฐ€ reverse process๋ฅผ ๊ฑฐ์ณ์„œ $rev(x)$๋ฅผ ์–ป๋Š”๋‹ค.
  2. 1๋ฒˆ ๊ณผ์ •์„ K๋ฒˆ ๋ฐ˜๋ณตํ•˜์—ฌ ${rev(x)_1,โ€ฆ,rev(x)_K}$๋ฅผ ์–ป๋Š”๋‹ค.
  3. ${rev(x)_1,โ€ฆ,rev(x)_K}$๋ฅผ classifier์— ํ†ต๊ณผ์‹œ์ผœ ๊ทธ ์ค‘์—์„œ ๊ฐ€์žฅ ๋งŽ์ด ๋‚˜์˜จ label์„ ์ตœ์ข… ์˜ˆ์ธก๊ฐ’์œผ๋กœ ์„ค์ •ํ•œ๋‹ค. ์ด๋ ‡๊ฒŒ ๊ฐ€์žฅ ๋งŽ์ด ๋‚˜์˜ค๋Š” label์„ ์ฐพ๋Š” ๊ณผ์ •์„ Majority Vote, ์ค„์—ฌ์„œ MV๋ผ๊ณ  ํ•œ๋‹ค.

๋˜ํ•œ DensePure์—

  1. Randomized Smoothing์„ ์ ์šฉํ•ด $L_2$-norm adversarial attack์„ ํšจ๊ณผ์ ์œผ๋กœ ๋ง‰๋Š” ๊ฒƒ์— ๋Œ€ํ•œ ์ฆ๋ช…๊ณผ
  2. Improved Denoising Diffusion Probablistic Models์— ์‚ฌ์šฉ๋œ Fast Sampling ๊ธฐ๋ฒ•์„ ์‚ฌ์šฉํ•ด ๋” ๋น ๋ฅด๊ฒŒ ์ƒ˜ํ”Œ๋งํ•˜๋Š” ๊ณผ์ •์— ๋Œ€ํ•ด์„œ ์„ค๋ช…ํ•˜๊ณ  ์žˆ์œผ๋‚˜ ์ค‘์š”ํ•œ ๋‚ด์šฉ์€ ์•„๋‹ˆ๋ฏ€๋กœ ์ƒ๋žตํ•˜๋„๋ก ํ•˜๊ฒ ๋‹ค.
    ๋‹ค์Œ์€ DensePure์˜ pipeline์— ๋Œ€ํ•ด ๋‚˜ํƒ€๋‚ธ figure์ด๋‹ค.

Experiments

์•„๋ž˜๋Š” ๋‹ค๋ฅธ baseline method์™€์˜ ์„ฑ๋Šฅ ๋น„๊ต๋ฅผ ํ†ตํ•ด ๋‚˜ํƒ€๋‚œ ๊ฒฐ๊ณผ์ด๋‹ค. ๋ฐ์ดํ„ฐ์…‹์€ CIFAR-10, ImageNet์„ ์‚ฌ์šฉํ•˜์˜€๋‹ค.

์—ฌ๊ธฐ์„œ $\epsilon$์€ adversarial attack์— ์˜ํ•ด perturbed๋œ ๋น„์œจ์ด๊ณ , ๊ฐ ์ •ํ™•๋„ ์ˆ˜์น˜ ์˜†์— ๊ด„ํ˜ธ๋Š” $\epsilon=0$ ์ผ ๋•Œ์˜ ์ •ํ™•๋„์ด๋‹ค. ์ฆ‰, standard accuracy ์ด๋‹ค. ๋˜ํ•œ off-the-shelf๋Š” ๋ชจ๋ธ์ด๋‚˜ classifier์˜ ๋ณ„๋„์˜ ํ•™์Šต์„ ์š”๊ตฌํ•˜์ง€ ์•Š๋Š” plug-and-play manner๋กœ ์ž‘๋™ํ•˜๋Š” method๋ฅผ ์ง€์นญํ•œ๋‹ค. ๊ฒฐ๊ณผํ‘œ๋ฅผ ๋ณด๋ฉด ์•Œ ์ˆ˜ ์žˆ๋“ฏ์ด ๋‹ค๋ฅธ ๋ฐฉ๋ฒ•๋“ค๊ณผ ๋น„๊ต๋ฅผ ํ–ˆ์„ ๋•Œ ๊ฑฐ์˜ ๋ชจ๋“  $\epsilon$์— ๋Œ€ํ•ด์„œ SOTA๋ฅผ ๋‹ฌ์„ฑํ•˜๊ณ  ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ๋‹ค.

๋˜ํ•œ DensePure๊ณผ ๋น„์Šทํ•˜๊ฒŒ DM์„ ์‚ฌ์šฉํ•œ ๋…ผ๋ฌธ์ธโ€™(certified!!) adversarial robustness for free!โ€™ ์™€์˜ ์„ฑ๋Šฅ ๋น„๊ต๋ฅผ ํ•˜๊ณ ์žˆ๋‹ค. ๋‹ค์Œ ๊ทธ๋ž˜ํ”„๋Š” CIFAR-10, ImageNet ๋ฐ์ดํ„ฐ์…‹์— ๋Œ€ํ•œ ๋น„๊ต ๊ฒฐ๊ณผ์ด๋‹ค.

๋ชจ๋“  noise scale $\sigma$์— ๋Œ€ํ•ด์„œ DensePure๊ฐ€ ๋” ์ข‹์€ ์„ฑ๋Šฅ์„ ๋ณด์ด๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ๋‹ค.

Ablation study

์•„๋ž˜ figure์€ Voting sample์˜ $K$๊ฐ’๊ณผ Fast sampling steps $b$์— ๋Œ€ํ•œ ablation study๋ฅผ ์ง„ํ–‰ํ•œ ๊ฒฐ๊ณผ์ด๋‹ค.

๊ฒฐ๊ณผํ‘œ๋ฅผ ํ†ตํ•ด ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์‚ฌ์‹ค์„ ์•Œ ์ˆ˜ ์žˆ๋‹ค.

  1. $K$๊ฐ’์ด ์ฆ๊ฐ€ํ•  ๋•Œ๋งˆ๋‹ค ์ •ํ™•๋„๊ฐ€ ์ฆ๊ฐ€ํ•œ๋‹ค.
  2. MV(Majority Vote)๊ฐ€ ์‚ฌ์šฉ๋˜๋ฉด sampling step์ด ์ฆ๊ฐ€ํ•  ๋•Œ๋งˆ๋‹ค ์ •ํ™•๋„๊ฐ€ ์ฆ๊ฐ€ํ•˜์ง€๋งŒ, MV(Majority Vote)๊ฐ€ ์‚ฌ์šฉ๋˜์ง€ ์•Š์œผ๋ฉด sampling step์ด ์ฆ๊ฐ€ํ•  ๋•Œ๋งˆ๋‹ค ์ •ํ™•๋„๊ฐ€ ๊ฐ์†Œํ•œ๋‹ค.

Limitations

MV์— ์‚ฌ์šฉ๋˜๋Š” reverse process ๋•Œ๋ฌธ์— time complexity๊ฐ€ ๋งค์šฐ ๋†’๊ฒŒ ๋‚˜ํƒ€๋‚œ๋‹ค. ํ•ด๋‹น ๋…ผ๋ฌธ์—์„œ๋Š” ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋ฅผ ๊ทน๋ณตํ•˜๊ธฐ ์œ„ํ•ด fast sampling ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ–ˆ์ง€๋งŒ ๋ฌธ์ œ๊ฐ€ ์™„์ „ํžˆ ํ•ด๊ฒฐ๋œ ๊ฒƒ์€ ์•„๋‹ˆ๊ธฐ ๋•Œ๋ฌธ์— ๋” ๋ฐœ์ „๋œ fast sampling ๋ฐฉ๋ฒ•์ด ํ•„์š”ํ•  ๊ฒƒ์ด๋‹ค.

Comments