NAWE (Neural-Assisted Watermark Extraction) combines an explicit signal-processing watermarking construction with a pretrained neural host predictor. A periodic, perceptually masked watermark carrier provides synchronization, Polar coding supplies redundancy, and denoising followed by subtraction extracts the embedded watermark. The denoiser remains frozen, without watermark-specific training. A one-factor-at-a-time study compares Wiener, BM3D, DRUNet, and GS-DRUNet host estimators. Comparisons with TrustMark, SSL Watermarking, PixelSeal, and WAM show NAWE's lowest geometric and photometric class BER and strong message recovery, while filtering and noise remain limitations consistent with the non-adaptive selection of the watermark extractor. The comparison retains the systems' different payloads and coding.
Figures & tables
Figure 1: NAWE signal path. Blue blocks perform analytical processing, green blocks carry image or message evidence, and orange denotes the frozen neural host predictor. The host determines the perceptual mask; the independent message and keyed synchronization bits determine the carrier. The receiver shares the key, nonce, layout, and decoder settings.
Figure 2: Composition of the 156-arm benchmark across 34 attack families. Inner sectors: geometry (44), photometric/color (26), compression/quantization (31), filtering/restoration/noise (25), and compound (30). Outer sectors show severity-arm counts.
Figure 3: OFAT sensitivity as ΔBER=BERarm−BERP0 (smaller is better). BERarm is the BER of a construction with a single control parameter changed and BERP0 that of the reference P0 under the set of attacks that is comprised of 34 arms from 7 families (reduced form of Figure 2 ), among which are JPEG, Gaussian blur and noise, resize, rotation, crop and flip on 100 images. BERP0=0.07 . Only one control parameter changes per row.
Figure 4: Family-balanced WER versus BER over the robustness matrix in Figure 2 , on logarithmic axes. Baseline curves sweep PSNR(x,y)∈[34,48] dB; NAWE markers show 48, 40, and 34 dB for protected-frame lengths L∈{32,64,256} , all at code rate 1/4 . Legend subscripts give the evaluated bit length.
Figure 5: Per method BER at 40 dB across the 34 attack families from Figure 2 benchmark. Each marker is μ BER across all arms of a single attack from X axis and upward whisker is σ . Dashed horizontal lines represent μ of each distortion class. NAWE has the lowest geometric and photometric class averages.