Page 53 - 2025S

46 UEC Int’l Mini-Conference No.54

Table 1: Robustness under common image process-
ing attacks

Attack Recovered BER Recognition
No modification 100.00% 0.00% Excellent
Superimposed figure 96.22% 0.03% Good
JPEG (QF=50) 37.70% 62.30% Poor
Gaussian (σ = 0.01) 21.10% 50.35% Poor
Contrast adjusted 46.97% 53.03% Poor
(a) Image with JPEG com- (b) Mark broken by com-
pression pression
• Localization accuracy: Within ±4 pixels of
the tampered region.

• PSNR: Infinite when extraction is perfect (no
attacks).

• SSIM: 0.998 when watermark occupies up to
50% of capacity.

• False positive rate: 1.2% in tamper localiza- (c) Image with Gaussian (d) Mark broken by Gaus-
tion. noise sian noise

(a) Original im- (c) Extracted re-
age (b) Mask gion
(e) Image with contrast ad- (f) Mark broken by contrast
Figure 7: Watermark extraction stage without tam- justment adjustment
pering (a–c).
Figure 9: System robustness under various attacks:
(a–b) JPEG compression, (c–d) Gaussian noise,
(e–f) Contrast adjustment.

the original image through a block-based approach,
LSB manipulation, and chaotic encryption.
(a) Tampered Given a watermarked image I , the procedure fol-
′
image (b) Tamper mask (c) Recovered lows:
Figure 8: Tamper detection and recovery stage (d–f).
1. Block division: I is partitioned into non-
′
overlapping 4×4 pixel blocks
3.3 Extraction stage 2. Bit extraction: Each block is analyzed to re-
trieve LSBs (excluding the seed pixel)
The extraction process in the proposed scheme en-
ables reversible recovery of hidden clinical informa- 3. Reconstruction: Extracted bits are decrypted
tion (watermark) embedded in medical images with- and reorganized to obtain the original water-
out data loss. This method preserves the integrity of mark