Light Field Messaging with Photographic Steganography

Post on 05-Nov-2021

5 views 0 download

Transcript of Light Field Messaging with Photographic Steganography

Pixel 2 & Samsung 2494SJ

Basler acA2040-90uc & Acer S240ML

Logitech c920 & Insignia NS-40D40SNA14

iPhone 8 & Acer Predator XB271HU

Basler acA1300-30uc & Dell 1707FPt

LFM without 𝑻(), frontal 49.96% 50.14% 50.05% 50.11% 50.04%LFM with 𝑻(), 45Β°(ours) 29.81% 15.23% 𝟏𝟎. 𝟐𝟐% 5.14% 10.01%LFM with 𝑻(), frontal (ours) 𝟏𝟎. πŸŽπŸ“% πŸ”. πŸ“πŸ–% 10.33% πŸ“. πŸŽπŸ•% πŸ’. πŸ–πŸ‘%

Camera-Display 1M Dataset:

1,000,000+ Images from 25 Camera-Display Pairs

Objective Function

minimize: 𝐿2|π’Šπ’„β€² βˆ’ π’Šπ’„| + 𝐿1|π’Šπ’Ž

β€² βˆ’ π’Šπ’Ž|subject to: 𝐸 π’Šπ’„, π’Šπ’Ž = π’Šπ’„

β€² coded image𝑇 π’Šπ’„

β€² = π’Šπ’„β€²β€² after transmission

𝑅 π’Šπ’„β€²β€² = π’Šπ’Ž

β€² recovered message

Original Images

Pixel 2 & Samsung

2494SJ

Basler acA2040-90uc

& Acer S240ML

Logitech c920 & Insignia

NS-40D40SNA14

iPhone 8 & Acer Predator

XB271HU

Basler acA1300-30uc & Dell 1707FPt

Light field messaging (LFM) transmits imperceptible, camera-readable messages with minimal bit error rate (BER).

Light Field Messaging with Photographic SteganographyEric Wengrowski and Kristin Dana

Rutgers University Department of Electrical and Computer Engineering, Steg AI

CNN Architecture

Sony CybershotDSC-RX100 & Lenovo Thinkpad X1 Carbon 3444-CUU

Sony CybershotDSC-RX100 & Apple MacbookPro 13-inch, Early 2011

Nikon Coolpix S6000 & Lenovo Thinkpad X1 Carbon 3444-CUU

Nikon Coolpix S6000 & Apple Macbook Pro, 13-inch Early 2011

DCT (Kamya, 2014), frontal 50.01% 50.13% 50.00% 49.95%Baluja (NeurIPS 2017), frontal 40.37% 37.15% 48.50% 48.83%LFM without 𝑇(), frontal 50.06% 49.95% 50.00% 50.00%LFM with 𝑇(), 45Β°(ours) 12.97% 15.59% 27.43% 25.81%LFM with 𝑻(), frontal (ours) πŸ—. πŸπŸ•% πŸ•. πŸ‘πŸ% 𝟐𝟎. πŸ’πŸ“% πŸπŸ•. πŸ“πŸ”%

Modeling Camera-Display Transfer

Function (CDTF) in Training

LFM Trained Without 𝑇() (πŸ’πŸ”. πŸ‘πŸ—% BER)

Coded Image π’Šπ’„β€² Residual (π’Šπ’„

β€² βˆ’ π’Šπ’„) Message π’Šπ’Žβ€²

LFM Trained With 𝑇() (𝟏. πŸπŸ•% BER)

Coded Image π’Šπ’„β€² Residual (π’Šπ’„

β€² βˆ’ π’Šπ’„) Message π’Šπ’Žβ€²

Perceptual Loss Ablation Study

Original πœ†π‘‡ = 0 πœ†π‘‡ = 0.001 πœ†π‘‡ = 0.01

Results: Robust to Exposure Settings and Viewing Angle

Over Exposed Normal Under Exposed

7.42% BER 0.78% BER 0.29% BER

30Β° Viewing Angle Homography 2.73% BER

45Β° Viewing Angle Homography 11.72% BER

π’Šπ’„

π’Šπ’Ž

π’Šπ’„β€²

π’Šπ’„β€²β€²

π’Šπ’Žβ€²

Compare to Baluja (NIPS 2017)

Results:

BER (Lower is

Better)

Compare to

Wengrowski et al.

(WACV 2016)

Original

Wengrowski

et al. (WACV

2016)

LFM (ours)

Results:

BER on New

Hardware (Lower

is Better)