Combating synthetic imagery in battlefield conditions with machine learning techniques

This work evaluates the efficacy of machine learning detection techniques against diffusion-generated synthetic satellite imagery under realistic battlefield degradation conditions. Two benchmark approaches are assessed: a contrastive language-image pre-training (CLIP) semantic embedding classifier and a NoisePrint convolutional neural network (CNN) noise-residual classifier. A novel hybrid detection framework which fuses semantic and forensic feature representations into a unified 576-dimensional embedding space. Synthetic imagery is generated using Stable Diffusion 3.5 and subjected to three levels of degradation, specifically clean, degraded, and extreme degraded conditions, simulating compression and corruption artifacts common to tactical intelligence, surveillance, and reconnaissance (ISR) environments. Results demonstrate that semantic features dominate detection performance. The CLIP-based classifier achieves near-perfect classification across all conditions (ROC-AUC ≥ 0.992). The NoisePrint classifier alone exhibits stronger robustness under moderate degradation (AUC = 0.968) but degrades substantially under extreme conditions (AUC = 0.666). The fused hybrid model achieves near-perfect AUC across all conditions, reaching 0.993 at extreme degradation, and improves calibration and confidence over single-modality approaches. These findings highlight the limitations of classical forensic noise methods on modern diffusion imagery and demonstrate the utility of multi-modal detection strategies for constrained ISR applications.