Reliable Evaluation of Adversarial Robustness with an Ensemble of Diverse Parameter free Attacks

1 minute read

1. Introduction

  • ์ง€๊ธˆ๊นŒ์ง€ ๋งŽ์€ adversarial defense ๋ฐฉ๋ฒ•์ด ๋‚˜ํƒ€๋‚ฌ์ง€๋งŒ ์ถ”ํ›„์— ๋“ฑ์žฅํ•œ ๋ฐœ์ „๋œ ๊ณต๊ฒฉ๋ฐฉ๋ฒ•์— ๋ฌด๋„ˆ์ง€๋Š” ๋ชจ์Šต๋“ค์ด ๋งŽ์ด ๋‚˜ํƒ€๋‚ฌ๋‹ค.

  • PGD๋Š” ๋ชจ๋ธ์˜ adversarial robustness๋ฅผ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ๋Š” ์ข‹์€ ๊ณต๊ฒฉ๋ฐฉ๋ฒ•์ด์˜€์ง€๋งŒ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ๋ฉด์—์„œ ์—ฌ๋Ÿฌ๊ฐ€์ง€ ๋ฌธ์ œ๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ๋‹ค.

  1. ๊ณ ์ •๋œ step size
  2. cross entropy loss
  • ํ•ด๋‹น ๋…ผ๋ฌธ์—์„œ๋Š” ์œ„์˜ ๋‘๊ฐ€์ง€ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” Auto-PGD๋ฅผ ์ œ์‹œํ•œ๋‹ค. Auto-PGD๋Š” PGD์™€๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์€ ๋ฉด์—์„œ ์ฐจ์ด๋ฅผ ๋ณด์ธ๋‹ค.
  1. step size๊ฐ€ adaptiveํ•˜๊ฒŒ ์„ ํƒ์ด ๋œ๋‹ค.
  2. cross entropy loss์™€๋Š” ๋ณ„๋„์˜ ์†์‹คํ•จ์ˆ˜๊ฐ€ ์‚ฌ์šฉ๋œ๋‹ค.
  3. ์กฐ์ ˆ ๊ฐ€๋Šฅํ•œ parameter๋กœ gradient step upadte์˜ ๋ฐ˜๋ณตํšŸ์ˆ˜๋งŒ์œผ๋กœ ์„ค์ •ํ•˜์—ฌ ์ฃผ์–ด์ง„ ์ž์›์— ์•Œ๋งž๊ฒŒ adpativeํ•œ ๊ณต๊ฒฉ์„ ํ•  ์ˆ˜ ์žˆ๋„๋ก ์„ค์ •ํ•˜์˜€๋‹ค.
  • ๋˜ํ•œ ๊ณต๊ฒฉ๋ฐฉ๋ฒ•์˜ ๋‹ค์–‘์„ฑ ๋ถ€์กฑ์€ ๋ชจ๋ธ์˜ robustness๋ฅผ ๊ณผ๋Œ€ํ‰๊ฐ€ ํ•˜๋„๋ก ๋งŒ๋“ค์—ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค์–‘ํ•œ ๊ณต๊ฒฉ๋ฐฉ๋ฒ•์„ ensembleํ•œ AutoAttack์„ ์ œ์‹œํ•œ๋‹ค.

2. Adversarial example and PGD

Adversarial example์€ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์ƒ์„ฑ๋œ๋‹ค.


$argmax_{k=1,โ€ฆ,K}\,g_k(z)\neq c$, $d(x_{orig}, z) \le \epsilon$

์—ฌ๊ธฐ์„œ $z$๋Š” adversarial sample, $c$๋Š” perturb ๋˜๊ธฐ ์ „์— ๋ถ„๋ฅ˜๋˜๋˜ ํด๋ž˜์Šค์ด๋‹ค.
์ตœ์ ์˜ $z$๋ฅผ ์ฐพ๊ธฐ ์œ„ํ•ด ๋‹ค์Œ๊ณผ ๊ฐ™์€ surrogate ํ•จ์ˆ˜๊ฐ€ ์‚ฌ์šฉ๋œ๋‹ค.


$max_{z \in D}L(g(z), c)$ such that $\gamma (x_{orig}, z) \le \epsilon, z \in D$

์—ฌ๊ธฐ์„œ $x$ ์™€ $z$์‚ฌ์ด์˜ ๊ฑฐ๋ฆฌ, ์ฆ‰ $\lVert z-x \rVert_p$๋ฅผ ์ตœ๋Œ€ํ™”ํ•˜๋Š” ๋ฐฉํ–ฅ์œผ๋กœ x๊ฐ€ ์›€์ง์—ฌ์•ผ ํ•˜๋ฏ€๋กœ ๋‹ค์Œ๊ณผ ๊ฐ™์€ Projected Gradient Descent(PGD)๊ฐ€ ์‚ฌ์šฉ๋œ๋‹ค.


$k = 1,โ€ฆ,N_{iter}$ as $x^{(k+1)} = P_s(x^{(k)}+\eta^{(k)}\nabla f(x^{(k)}))$

์—ฌ๊ธฐ์„œ $f:R^d \to R, S \subset R^d$์ด๊ณ  $P_s$๋Š” $S$์— ๋Œ€ํ•œ ์‚ฌ์˜, $L$๋กœ๋Š” cross entroy loss๊ฐ€ ์‚ฌ์šฉ๋œ๋‹ค.

3. Auto-PGD

PGD๋Š” 1. ๊ณ ์ •๋œ step size์™€ 2. ์ œ๊ณต๋œ ์ž์›์— ์˜ํ–ฅ์„ ๋งŽ์ด ๋ฐ›์Œ ๋ฉด์—์„œ ์—ฌ๋Ÿฌ ๋ถˆ์•ˆ์ •ํ•จ์„ ๊ฐ€์ง€๊ณ  ์žˆ๋‹ค.
์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด Auto-PGD๋Š” $N_{iter}$์„ 1. ์ข‹์€ ์‹œ์ž‘ ์ง€์ ์„ ์ฐพ๋Š” exploration phase, 2. ์ง€๊ธˆ๊นŒ์ง€ ์ถ•์ ๋œ ์ •๋ณด๋ฅผ ์ตœ๋Œ€ํ™” ์‹œํ‚ค๋Š” exploitation phase๋กœ ๋‚˜๋ˆˆ๋‹ค. ์—ฌ๊ธฐ์„œ exploration phase์—์„œ๋Š” ํฐ step size์œผ๋กœ ๋น ๋ฅด๊ฒŒ ์‹œ์ž‘ ์ง€์ ์„ ์ฐพ๊ณ  exploitation phase์—์„œ๋Š” ์ž‘์€ step size๋กœ $f$๋ฅผ ์ตœ๋Œ€ํ™” ์‹œํ‚จ๋‹ค. ์ด๋ ‡๊ฒŒ step size์˜ ํฌ๊ธฐ๋ฅผ ๋‹ค์–‘ํ•˜๊ฒŒ ์กฐ์ ˆํ•˜๋Š” ๊ฒƒ์ด Auto-PGD์˜ ํ•ต์‹ฌ์ด๋ผ๊ณ  ํ•  ์ˆ˜ ์žˆ๋‹ค. ๋จผ์ € Auto-PGD์˜ ์ „์ฒด์ ์ธ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์‚ดํŽด๋ณด๋ฉด ๋‹ค์Œ๊ณผ ๊ฐ™๋‹ค.

์œ„ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์ž์„ธํžˆ ํ•˜๋‚˜์”ฉ ์‚ดํŽด๋ณด์ž.

1. Gradient step: ๋จผ์ € Auto-PGD์˜ gradient step์€ PGD์˜ gradient step์— momentum term์„ ์ถ”๊ฐ€ํ•œ๋‹ค.

2. Step size selection: ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์กฐ๊ฑด์„ ๋งŒ์กฑํ•˜๋ฉด step size๋ฅผ ๋ฐ˜์œผ๋กœ ๋‚˜๋ˆˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  step size์— ํ•ด๋‹นํ•˜๋Š” ์ฒดํฌํฌ์ธํŠธ๋ฅผ $w_0 = 0, w_1,โ€ฆ,w_n$๋กœ ์ง€์ •ํ•œ๋‹ค.

๋งŒ์•ฝ ์ฒดํฌํฌ์ธํŠธ $w_j$๊ฐ€ ๋ฐ˜์œผ๋กœ ๋‚˜๋ˆ„์–ด์ง€๋ฉด $x^{(w_j+1)}$๋ฅผ $x_max$๋กœ ์„ค์ •ํ•˜๊ณ  $f_max$์—์„œ๋ถ€ํ„ฐ ๋‹ค์‹œ ์‹œ์ž‘ํ•ด exploitation phase๋กœ $f$๋ฅผ ์ตœ๋Œ€ํ™” ์‹œํ‚จ๋‹ค. ์ด ๋•Œ ์ฒดํฌํฌ์ธํŠธ $w_j$์—์„œ๋งŒ step size๋ฅผ ๊ฐ์†Œ์‹œํ‚ค๋Š” ๊ฒƒ ๋ฟ๋งŒ์ด ์•„๋‹Œ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์ง€์†์ ์ธ ๊ฐ์†Œ๋ฅผ ํ†ตํ•ด localized search๋ฅผ ํ•˜๊ฒŒ ๋งŒ๋“ ๋‹ค.

$p_{j+1}-p_j$๊ฐ€ 0.03๋งŒํผ ๊ฐ์†Œํ•˜๋˜, ์ตœ์†Œ 0.06๊ฐ€ ๋˜๋„๋ก ๋งŒ๋“ค์—ˆ๋‹ค.

Comments