TRACER: Transformer-based Digital Forensics Framework for Detecting and Reconstructing Adversarial Attacks in Vision-Language Models
International Journal of Innovations in Science and Technology · 30.09.2026 · 8 (7) · 3084
DOI: 10.33411/ijist/2071 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Innovations in Science and Technology, online erschienen am 30.09.2026. Verfasst von Iqra Yasmeen und 5 weiteren Autor:innen. Die Arbeit nennt 51 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
In the real world, Vision-Language Models (VLMs) have become a more common sight in security-sensitive settings, and it is possible to introduce subtle, and often visually undetectable, changes to their inputs that can trick the models into producing incorrect multimodal reasoning, especially in the context of digital forensics. In this article, we present TRACER (Transformer-based Reconstruction and Attack Characterization to Recover Evidence): a system capable of identifying, characterizing and reconstructing adversarial attacks against VLMs without relying on without a priori information, watermarking or access to the source attack process. The central concept is to use transformer architectures' intrinsic attention dynamics to detect adversarial behavior by monitoring irregularities in attention distributions using entropy. A gradient-based attribution model is then used to identify the affected regions to create fine-grained forensic maps of the hostile trigger zones, and a diffusion-based reconstruction module uses both the attribution map and adversarial signal in a single dual-stream design to reconstruct a semantically consistent, clean estimate of the input for end-to-end forensic analysis. The multimodal experimental evaluation shows that TRACER can deliver high detection accuracy while maintaining good reconstruction fidelity, outperforming previous baselines in adversarial detection, spatial attribution and interpretability. The results show the feasibility of passive, attention-based forensic analysis as a scalable and reliable security layer for next-generation multimodal AI systems. TRACER can accurately detect FGSM and PGD perturbations (94.2-99.1% accuracy, 0.962-0.995 AUC) and localise the perturbation at IoU 0.71-0.87 with respect to the ground-truth perturbation mask or up to 0.91 with respect to the exact mask obtained by localized patch attacks, and reconstruct the input at 27.4-32.8 dB PSNR and 0.86-0.94 SSIM, and recover the downstream task accuracy to 89.4% on average after the attack. It achieves an accuracy advantage of 8.6 percentage points.cover the statistical-feature detector, 0.19 IoU over Grad-CAM saliency and 4.2 dB (all paired t-tests p < 0.001) over a diffusion purification approach with no steering, and adds only 38 ms of latency and 14% of compute to the inference path while mislabelling 3.1% of clean inputs. Results are averaged over 5 different runs for each attack/perturbation-budget combinations (FGSM and PGD with epsilon = {2/255, 4/255, 8/255}) and the standard deviation across the different runs is no more than +/-0.5 percentage points for detection accuracy and +/-0.3 dB for reconstruction PSNR.
Zitieren
APA
Yasmeen, I., Firdous, S., Ahsan, T., Malik, A. H., Ashraf, A., & Asad, M. (2026). TRACER: Transformer-based Digital Forensics Framework for Detecting and Reconstructing Adversarial Attacks in Vision-Language Models. International Journal of Innovations in Science and Technology, 8(7), 3084. https://doi.org/10.33411/ijist/2071
Vancouver
Yasmeen I, Firdous S, Ahsan T, Malik AH, Ashraf A, Asad M. TRACER: Transformer-based Digital Forensics Framework for Detecting and Reconstructing Adversarial Attacks in Vision-Language Models. International Journal of Innovations in Science and Technology. 2026;8(7):3084. doi:10.33411/ijist/2071
BibTeX
@article{Yasmeen2026,
author = {Yasmeen, Iqra and Firdous, Saba and Ahsan, Talha and Malik, Abdul Haseeb and Ashraf, Asiya and Asad, Muhammad},
title = {TRACER: Transformer-based Digital Forensics Framework for Detecting and Reconstructing Adversarial Attacks in Vision-Language Models},
journal = {International Journal of Innovations in Science and Technology},
year = {2026},
volume = {8},
number = {7},
pages = {3084},
doi = {10.33411/ijist/2071}
}
EUVIMED