Presentation Schedule
Pioneering Hybrid Multi-scale Forgery Detection: AI-generated Image Discrimination at Early Stages via Adaptive Frequency-CLIP Fusion (105810)
Monday, 9 November 2026 16:15
Session: Poster Session
Room: Atrium (1F)
Presentation Type:Poster Presentation
Detecting image forgery in the early stages of production-before post-processing is mature-poses unprecedented risks to media integrity, academia and national security infrastructures, as nascent AI models, including propagation-based generative adversarial networks, produce nearly indistinguishable synthetic content. Contemporary detection frameworks exhibit significant performance degradation when faced with immature production outputs characterized by minimal post-processing artifacts and subtle statistical anomalies. This research shows and introduces EFN (EarlyForgeNet), A methodological framework of standard accuracy Q1 that synergistically integrates multilocal intrinsic dimension analysis (multiLID) with FAA (Forgery-Aware Adapter) and LGA (Language-Guided Alignment) mechanisms derived from transformer architectures, which are significantly improved through quantum-inspired wavelet transforms and achieve sub-1.1% error rates on immature synthetic datasets. The proposed method was subjected to rigorous evaluation across cross-domain benchmarks including CIFAR-10 synthetic subsets and early generations of ProGAN, and demonstrated 98.8% (AUC) area under the receiver operating characteristic curve, outperforming contemporary baseline approaches by 14-22.1 % in low-artifact scenarios. According to a systematic review conducted by the Office of Research Integrity, international epidemiological data shows a 48% increase in undetected academic image fraud between 2020 and 2024. The framework includes future-proof mechanisms including federated learning architectures for real-time deployment scenarios and blockchain-based provenance verification. Statistical validation via paired t-tests confirms significance at p<0.001 levels, with F1 scores reaching 0.98 under heterogeneous testing conditions. This research establishes new paradigms for early-stage detection of synthetic content and directly addresses critical gaps in digital forensics where conventional post-processing detection methods do not demonstrate sufficient sensitivity.
Authors:
Ata ul kareem, Allama Tabataba'i University, Iran
Jazba Nudrat, Nusrat Jahan College, Pakistan
About the Presenter(s)
Ata ul Kareem currently Phd student at Allama Tabataba'i Univesity Tehran-Iran as a foreign student.
See this presentation on the full schedule – Monday Schedule
Conference Comments & Feedback
Place a comment using your LinkedIn profile[wplinkpress_comments]
Share this Presentation
