Stealthy Porn: Understanding Real-World Adversarial Images for - - PowerPoint PPT Presentation

▶
stealthy porn understanding real world adversarial images
SMART_READER_LITE
LIVE PREVIEW

Stealthy Porn: Understanding Real-World Adversarial Images for - - PowerPoint PPT Presentation

Oct 03, 2022 191 likes •495 views

Stealthy Porn: Understanding Real-World Adversarial Images for Illicit Online Promotion Yuan, Di Tang , Xiaojing Liao, XiaoFeng Wang, Xuan Feng, Kan Yi Chen, Menghan Sun, Haoran Lu, Kehuan Zhang A C ASE IN THE W ILD Sexual stream t.cn

slide-1
SLIDE 1

Stealthy Porn: Understanding Real-World Adversarial Images for Illicit Online Promotion

Kan Yuan, Di Tang, Xiaojing Liao, XiaoFeng Wang, Xuan Feng, Yi Chen, Menghan Sun, Haoran Lu, Kehuan Zhang

slide-2
SLIDE 2

A CASE IN THE WILD

Sexual stream t.cn

Sina’s URL shortener service

p*.z*.s*.com

SoHu redirector

s*.top landing domain Illicit content detector

Evade

slide-3
SLIDE 3

I am so attractive, and always on photos to catch the eyes.

I am yovr need

Earn your money

  • ADVERSARIAL IMAGES IN THE REAL-WORLD

APPIs

Adversarial Promotional Porn Images

slide-4
SLIDE 4

APPIS ARE DIFFERENT

  • small noise
  • nearly indistinguishable

Adversarial Examples Corrupted Images

  • large noise
  • meaningless

I am yovr need

slide-5
SLIDE 5

I am yovr need

APPIS COMPOSITION

Noise Promotional information Abnormal shape Transparent box Color strips

slide-6
SLIDE 6

Two common characters:

MALÈNA: FINDING STEALTHY PORN

I am yovr need

✤ Promotional content. ✤ less obfuscated explicit content

slide-7
SLIDE 7

I

image type promotional content explicit content evasiveness

MALÈNA: FINDING STEALTHY PORN

APPIs

slide-8
SLIDE 8

I

image type promotional content explicit content evasiveness

MALÈNA: FINDING STEALTHY PORN

APPIs

slide-9
SLIDE 9

I

image type promotional content explicit content evasiveness

MALÈNA: FINDING STEALTHY PORN

APPIs

slide-10
SLIDE 10

MALÈNA: PROMOTIONAL CONTENT IDENTIFIER

Text: PiexlLink QRcode: ZBar, ZXing, BoofCV WeChat

slide-11
SLIDE 11

MALÈNA: PROMOTIONAL CONTENT IDENTIFIER

Text: QRcode: PiexlLink

slide-12
SLIDE 12

I

image type promotional content explicit content evasiveness

MALÈNA: FINDING STEALTHY PORN

APPIs

slide-13
SLIDE 13

MALÈNA: REGIONAL EXPLICIT CONTENT DETECTOR

slide-14
SLIDE 14

MALÈNA: REGIONAL EXPLICIT CONTENT DETECTOR

slide-15
SLIDE 15

MALÈNA: REGIONAL EXPLICIT CONTENT DETECTOR

ResNet-50 Explicit

slide-16
SLIDE 16

I

image type promotional content explicit content evasiveness

MALÈNA: FINDING STEALTHY PORN

APPIs

slide-17
SLIDE 17

MALÈNA: PERFORMANCE

  • Performance: 91% precision , 85% recall
  • Result: 4,353/6,163 APPIs , from 4M images, 76K posts (Baidu

Tieba, Weibo)

slide-18
SLIDE 18

MEASUREMENT

✤ Visual pattern. ✤ Promotional content. ✤ Distribution channels.

slide-19
SLIDE 19

Original Blur Noise Rotation Occlusion Texture Color manipulation Transparentization

MEASUREMENT: VISUAL PATTERN

slide-20
SLIDE 20
  • Rotation

➤ 45 and 135 degrees are effective

MEASUREMENT: VISUAL PATTERN

slide-21
SLIDE 21
  • Rotation

➤ 45 and 135 degrees are effective

  • Noising

➤ Less noising is enough

MEASUREMENT: VISUAL PATTERN

slide-22
SLIDE 22
  • Rotation

➤ 45 and 135 degrees are effective

  • Noising

➤ Less noising is enough

  • Color-manipulation

➤ Green is evasive colour

MEASUREMENT: VISUAL PATTERN

slide-23
SLIDE 23

MEASUREMENT: PROMOTIONAL CONTENT

slide-24
SLIDE 24

MEASUREMENT: PROMOTIONAL CONTENT

slide-25
SLIDE 25

MEASUREMENT: PROMOTIONAL CONTENT

Reuse of promotional content

232 / 612 (37%)

Reuse of explicit content

3981 / 4353 (91%)

slide-26
SLIDE 26

MEASUREMENT: PROMOTIONAL CONTENT

Reuse of promotional content

232 / 612 (37%)

Reuse of explicit content

3981 / 4353 (91%)

slide-27
SLIDE 27

MEASUREMENT: DISTRIBUTION CHANNELS

Compromised accounts: rarely post comment only on hot microblog Dedicated accounts: > 30 posts/day with meaningless sentences

0% 15% 30% 45% 60% Compromised Dedicated

51% 24%

slide-28
SLIDE 28

LESSION LEARNED

Visual pattern. Harden current models. Promotional content. Regularize promotion channel. Distribution channels. Secure accounts.

slide-29
SLIDE 29

✤ APPIs are prevalent ✤ Understanding criminal goal and ecosystem behind

adversarial images

✤ Hardening machine learning model against APPI attack

deserves further studies

TAKE-AWAYS

slide-30
SLIDE 30

THANK YOU!