Work

NRGA-Net: Noise-residual Guided Attention Network for Local Manipulation Detection and Localization in Satellite Imagery

Year: 2026
Type: article
Abstract: Modern satellite imagery has become vulnerable to numerous manipulation threats due to the rapid progress of generative models, and forensic analysis is important to detect malicious edits and to secure the critical decisions that rely on remote-sensing data.Though, the purpose of this study is driven by the security issues that can be witnessed in satellite-based decision making, the presented framework is tested as a unified forgery localization and detection system with a reference to the benchmark forgery datasets.The suggested model combines a dilated DenseNet-201 semantic encoder with a dual-domain forensic lane that fuses a fixed-filter noise-residual stream and a learnable frequency-residual encoder through asymmetric cross-frequency attention, while a spectral-edge stream recovers boundaries in the Fourier domain, and the new Fake-LocalDiff benchmark adds prompt-driven latent-diffusion local replacements to the existing inpainting benchmarks.The proposed model is implemented and evaluated as a supervised forgery detector using labeled data, and it is evaluated on Fake-Vaihingen, Fake-LoveDA, and Fake-LocalDiff under an immutable train, calibration, and final-test protocol to analyze its effectiveness in localizing forged and authentic regions and in detecting manipulated images.Because each generative model imprints its own characteristics, features, and fingerprint on the forged images, the model is trained jointly on the three datasets (LaMa inpainting, RePaint, and latent diffusion), and the leave-one-family-out study shows that omitting one family from training weakens the model and prevents it from capturing the artifacts of the unseen generator.The performance on the independent final-test set is astounding, with a 97.80% pooled forged-class IoU, a Dice = F1 of 98.89%, a precision of 98.31%, a recall of 99.48%, and a 99.22% detection accuracy.The applicability of the proposed model to multispectral imagery and to unseen generator families is considered indicative, and future work will focus on contrastive validation of the provenance hash and on improving cross-generator transfer.The experimental analysis is done based on the three benchmarks, albeit under the incentive of the satellite-imagery security issues, thus the findings provide insights into the performance of the unified supervised framework on the benchmark protocols.
Language: English
Open Access status: hybrid
OpenAlex ID: W7235645502
FWCI:
Cites: 37
Cited by:
Related to: 10