Full text
| |
| Author(s): |
João Phillipe Cardenuto
Total Authors: 1
|
| Document type: | Doctoral Thesis |
| Press: | Campinas, SP. |
| Institution: | Universidade Estadual de Campinas (UNICAMP). Instituto de Computação |
| Defense date: | 2024-11-19 |
| Examining board members: |
Anderson de Rezende Rocha;
Renan Moritz Varnier Rodrigues Almeida;
André Carlos Ponce de Leon Ferreira de Carvalho;
Nina Sumiko Tomita Hirata;
Sérgio Luiz Monteiro Salles Filho
|
| Advisor: | Anderson de Rezende Rocha; Daniel Henriques Moreira |
| Abstract | |
Science has been reporting increasing misconduct cases, especially involving biomedical research. The scenario became critical when social media started to use fraudulent papers to propel misinformation. Despite all prior efforts to detect text fraud (e.g., plagiarism detection), there is a lack of studies addressing scientific image misconduct. Besides that, there is a daunting trend of problematic images published in scientific articles. Image misconduct is evolving from simple inappropriate edits and reuses to systematic manipulations involving potentially illegal organizations known as paper mills. Furthermore, with the advance of generative artificial intelligence (AI), it is expected that sooner or later, such organizations will fully synthesize scientific images –and even the entire article–, polluting the literature with fake data. In turn, the scientific community has been responding sluggishly, and only some trustworthy and effective solutions have been proposed, even for non-sophisticated image manipulation detection. Motivated by this intriguing scenario, this research investigates scientific integrity through a forensic lens. Due to the lack of prior forensic research, we organized and pointed to the main computational challenges regarding scientific image analysis, which we act toward mitigation. As a result, we proposed automated image analysis, document and image provenance analysis, and AI-image detectors focused on the biomedical area to assist integrity offices and journal editors during their decision-making. The proposed methods speed up the integrity analysis by using an end-to-end workflow that inputs a collection of PDF documents (e.g., a set of scientific articles), ending by highlighting suspicious images and documents that merit further attention. Following the recommendation of integrity offices, the proposed solutions use explainable steps with easy-to-interpret results for the analysts. In addition to the forensic methods, datasets, and benchmarks created—all open-source and freely available—this work also discusses the standard of biomedical images as scientific evidence and points to their prevalent problems, aiming to foster possible requirements for publishing trustworthy scientific images. This thesis results from collaborating with scientific integrity researchers and an international digital forensic team to mitigate the current threats to science and foster effective computational methods to improve research integrity (AU) | |
| FAPESP's process: | 20/02211-2 - Filtragem e Análise de Proveniência |
| Grantee: | João Phillipe Cardenuto |
| Support Opportunities: | Scholarships in Brazil - Doctorate (Direct) |