Physically-Guided Optical Inversion Enable Non-Contact Side-Channel Attack on Isolated Screens
Authors: Zhiwen Zheng, Yuheng Qiao, Xiaoshuai Zhang, Zhao Huang, Tao Zhang, Huiyu Zhou, Shaowei Jiang, Jin Liu, +2 more
Organizations: Hangzhou Dianzi University, Hangzhou 310018, China · Weidian Corporation, Beijing 100015, China · Zhejiang Provincial Key Laboratory of Low Altitude Ubiquitous Networking Technology, Hangzhou Dianzi University, Hangzhou, 310018, China · Faculty of Information Science and Engineering, Ocean University of China, Qingdao 266404, China · School of Natural and Computing Sciences, University of Aberdee, Aberdeen AB24 3UE, Great Britain · School of Computing and Mathematical Sciences, University of Leicester, Leicester LE1 7RH, UK · Johns Hopkins University, Baltimore, MD 21218, USA
Noncontact exfiltration of electronic screen content poses a security challenge, with side-channel incursions as the principal vector. We introduce an optical projection side-channel paradigm that confronts two core instabilities: (i) the near-singular Jacobian spectrum of projection mapping breaches Hadamard stability, rendering inversion hypersensitive to perturbations; (ii) irreversible compression in light transport obliterates global semantic cues, magnifying reconstruction ambiguity. Exploiting passive speckle patterns formed by diffuse reflection, our Irradiance Robust Radiometric Inversion Network (IR4Net) fuses a Physically Regularized Irradiance Approximation (PRIrr-Approximation), which embeds the radiative transfer equation in a learnable optimizer, with a contour-to-detail cross-scale reconstruction mechanism that arrests noise propagation. Moreover, an Irreversibility Constrained Semantic Reprojection (ICSR) module reinstates lost global structure through context-driven semantic mapping. Evaluated across four scene categories, IR4Net achieves fidelity beyond competing neural approaches while retaining resilience to illumination perturbations.
Figures & tables
Figure 1: In the figure, (a), (b), and (c) correspond to the rendered scene, schematic diagram, and real-world scene respectively.
Figure 2: Overall architecture of IR 4 Net, comprising the PRIrr‑Approximation and ICSR modules. The multi-scale frequency separation module, a key component of ICSR, is implemented via concatenation.
Methods
Source
ReSh-WebSight
ReSh-Password
ReSh-Screen
ReSh-Char
PSNR↑
RMSE↓
SSIM↑
PSNR↑
RMSE↓
SSIM↑
PSNR↑
RMSE↓
SSIM↑
PSNR↑
RMSE↓
SSIM↑
HVI-CIDNet
CVPR,25
18.940
33.837
0.792
13.024
57.269
0.858
21.027
26.686
0.708
15.720
44.843
0.692
DarkIR
CVPR,25
19.234
32.587
0.779
13.580
53.883
0.855
21.609
25.215
0.706
16.861
39.011
0.709
AST
CVPR,24
19.502
31.026
0.787
14.022
51.199
0.832
21.574
24.823
0.673
16.909
38.515
0.709
ConvIR
CVPR,24
19.573
30.678
0.799
14.779
47.077
0.867
22.010
23.718
0.731
16.678
39.569
0.707
C2PNet
CVPR,23
15.641
52.514
0.769
11.209
70.458
0.813
16.428
44.883
0.552
15.278
46.885
0.666
Table 1: Quantitative comparison of IR 4 Net against reconstruction-centric methods (HVI-CIDNet Yan et al. (2025) ,DarkIR Feijoo et al. (2025) ,AST Zhou et al. (2024) ,ConVIR Cui et al. (2024) ,C2PNet Zheng et al. (2023) ,Uformer Wang et al. (2022b) , UNet Ronneberger et al. (2015) ) and generation-centric methods (BicycleGAN Zhu et al. (2017b) ,Divco Liu et al. (2021) ,pix2pix Isola et al. (2017) ,CycleGAN Zhu et al. (2017a) ) on four benchmarks (ReSh-WebSight, ReSh-Password, ReSh-Screen, ReSh-Chart) under identical data splits and optimization.
Figure 3: Visual comparison of IR 4 Net and baseline methods on four datasets. Our model yields visually more faithful restorations across various scenes.
Metric
Chart
Screen
WebSight
OURS
ADMM
NAG
Heavyball
OURS
ADMM
NAG
Heavyball
OURS
ADMM
NAG
Heavyball
PSNR ↑
17.363
17.180
17.214
17.192
25.812
25.155
25.090
25.077
20.708
20.707
20.621
20.533
RMSE ↓
36.748
37.367
37.308
37.447
16.531
17.680
17.672
17.754
26.719
27.024
27.349
27.629
SSIM ↑
0.731
0.725
0.724
0.724
0.817
0.806
0.808
0.808
0.820
0.808
0.808
0.807
LPIPS ↓
0.431
0.468
0.465
0.462
0.216
0.232
0.235
0.231
0.282
0.299
0.299
0.300
Table 2: Performance comparison of four iterative schemes across three datasets.
Model
0
25
50
75
100
125
150
175
200
225
250
275
300
OURS
25.812
25.726
25.702
25.634
25.533
25.288
24.990
24.537
24.016
23.220
22.306
20.983
19.136
UNet
20.195
6.302
6.461
6.121
6.546
6.104
6.301
7.015
6.089
6.686
6.257
6.144
6.412
C2PNet
16.428
15.913
14.951
13.452
12.311
11.520
11.195
10.948
10.542
10.039
9.666
9.370
9.144
DarkIR
21.609
21.360
20.660
19.288
17.423
15.355
13.672
12.481
11.491
10.614
10.077
9.686
9.424
CIDNet
21.027
20.808
20.337
19.366
18.118
16.576
15.131
13.853
12.791
11.739
10.913
10.192
9.745
ConvIR
22.010
21.824
21.480
20.601
19.223
17.698
16.215
14.885
13.662
12.537
11.707
10.991
10.435
Table 3: PSNR comparison of different models under screen brightness reductions (in nits).
Figure 4: As screen brightness decreases on the ReSh‑Screen dataset, our model’s PSNR degrades significantly less than that of other methods.
Orbital motion
Camera rotation
Camera–wall distance
Pose (horizontal,vertical)
PSNR
SSIM
LPIPS
FID
Angle (pitch,yaw,roll)
PSNR
SSIM
LPIPS
FID
Dist. (m)
PSNR
SSIM
LPIPS
FID
(0,5)
25.046
0.801
0.226
1.047
(0, 0, 0)
25.812
0.817
0.216
0.967
2
25.810
0.817
0.216
0.970
(0,15)
22.814
0.749
0.263
0.953
(2, 0, 0)
25.586
0.814
0.219
0.995
3
25.791
0.816
0.217
0.967
(0,-5)
25.267
0.807
0.224
0.997
(5, 3, 2)
24.508
0.793
0.234
1.057
4
25.764
0.816
0.217
0.975
(10,0)
22.901
0.740
0.263
0.899
(8, 5, 4)
21.470
0.719
0.285
0.996
5
25.690
0.815
0.218
0.980
(15,0)
21.275
0.705
0.294
0.986
(0, -10, -3)
19.906
0.697
0.308
1.006
6
25.541
0.813
0.219
0.990
Table 4: Summary of IR 4 Net performance under three geometric perturbations: (1) orbital motion, (2) camera rotation, and (3) camera–wall distance variation. Each block lists five representative conditions from the full experiments.
Figure 5: Experimental setups for camera motion. (a) Camera orbital motion: (a1) motion along a horizontal arc parallel to the ground; (a2) motion along a vertical arc perpendicular to the ground (side view). (b) Camera rotation: (b1) tilting parallel to the projection wall; (b2) tilting perpendicular to the ground (side view); (b3) tilting perpendicular to the projection plane (top view). (c) Camera–wall distance: camera translating along the direction normal to the projection plane.
MOE Key Lab of ICSP, IMIS Lab of Anhui, Anhui Provincial Key Lab of Multimodal Cognitive Computation, Zenmorn-AHU AI Joint Lab, School of Computer Science and Technology, Anhui University, Hefei 230601, China
Electrical and Computer Engineering (ECE) Program Computer, Electrical, and Mathematical Science and Engineering (CEMSE) Division King Abdullah University of Science and Technology (KAUST) Thuwal, 23955-6900, Saudi Arabia