The rapid progress of generative AI, spanning GANs and diffusion models, has enabled the creation of highly realistic synthetic images, which are increasingly misused to spread propaganda and misinformation. Existing attribution techniques often assume access to the generator or require large numbers of samples, conditions rarely met in practice. We introduce a few-shot attribution framework based on tiny autoencoders that share a single decoder pretrained on natural images and learn lightweight, model-specific encoders from only a handful of samples. It yields a compact, modular solution that (i) consistently outperforms CNN/Transformer baselines, while remaining competitive with respect to state-of-the-art methods; (ii) can be easily extended to new candidate generators; (iii) remains robust under common post-processing (e.g., JPEG compression and resizing); and (iv) benefits from different entropy-based approaches which improve the patch selection criterion. We further show that the same reconstruction-error features naturally support joint detection and source attribution, making this approach practical for real-world forensic pipelines. We also validate its lightweight design through a browser-based implementation that runs locally on CPU, supporting privacy-preserving forensic analysis in resource-constrained environments.

Tiny autoencoders for few-shot source attribution of AI-generated images / Bindini L., Bertazzini G., Baracchi D., Shullani D., Frasconi P., Piva A.. - In: EURASIP JOURNAL ON INFORMATION SECURITY. - ISSN 2510-523X. - ELETTRONICO. - 2026:(2026), pp. 19.0-19.0. [10.1186/s13635-026-00240-y]

Tiny autoencoders for few-shot source attribution of AI-generated images

Bindini L.
;
Bertazzini G.;Baracchi D.;Shullani D.;Frasconi P.;Piva A.
2026

Abstract

The rapid progress of generative AI, spanning GANs and diffusion models, has enabled the creation of highly realistic synthetic images, which are increasingly misused to spread propaganda and misinformation. Existing attribution techniques often assume access to the generator or require large numbers of samples, conditions rarely met in practice. We introduce a few-shot attribution framework based on tiny autoencoders that share a single decoder pretrained on natural images and learn lightweight, model-specific encoders from only a handful of samples. It yields a compact, modular solution that (i) consistently outperforms CNN/Transformer baselines, while remaining competitive with respect to state-of-the-art methods; (ii) can be easily extended to new candidate generators; (iii) remains robust under common post-processing (e.g., JPEG compression and resizing); and (iv) benefits from different entropy-based approaches which improve the patch selection criterion. We further show that the same reconstruction-error features naturally support joint detection and source attribution, making this approach practical for real-world forensic pipelines. We also validate its lightweight design through a browser-based implementation that runs locally on CPU, supporting privacy-preserving forensic analysis in resource-constrained environments.
2026
2026
0
0
Bindini L.; Bertazzini G.; Baracchi D.; Shullani D.; Frasconi P.; Piva A.
File in questo prodotto:
File Dimensione Formato  
Tiny autoencoders for few-shot source attribution of AI-generated images.pdf

accesso aperto

Tipologia: Pdf editoriale (Version of record)
Licenza: Creative commons
Dimensione 2.12 MB
Formato Adobe PDF
2.12 MB Adobe PDF

I documenti in FLORE sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificatore per citare o creare un link a questa risorsa: https://hdl.handle.net/2158/1500391
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus ND
  • ???jsp.display-item.citation.isi??? ND
social impact