{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/exploring-multi-modal-fusion-for-image","title":"MMFusion: Combining Image Forensic Filters for Visual Manipulation Detection and Localization","arxiv_id":"2312.01790","date":"2023-12-04","proceeding":null,"authors":["Kostas Triaridis","Konstantinos Tsigos","Vasileios Mezaris"],"abstract":"Recent image manipulation localization and detection techniques typically leverage forensic artifacts and traces that are produced by a noise-sensitive filter, such as SRM or Bayar convolution. In this paper, we showcase that different filters commonly used in such approaches excel at unveiling different types of manipulations and provide complementary forensic traces. Thus, we explore ways of combining the outputs of such filters to leverage the complementary nature of the produced artifacts for performing image manipulation localization and detection (IMLD). We assess two distinct combination methods: one that produces independent features from each forensic filter and then fuses them (this is referred to as late fusion) and one that performs early mixing of different modal outputs and produces combined features (this is referred to as early fusion). We use the latter as a feature encoding mechanism, accompanied by a new decoding mechanism that encompasses feature re-weighting, for formulating the proposed MMFusion architecture. We demonstrate that MMFusion achieves competitive performance for both image manipulation localization and detection, outperforming state-of-the-art models across several image and video datasets. We also investigate further the contribution of each forensic filter within MMFusion for addressing different types of manipulations, building on recent AI explainability measures.","url_abs":"https://arxiv.org/abs/2312.01790v2","url_pdf":"https://arxiv.org/pdf/2312.01790v2.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"exploring-multi-modal-fusion-for-image","repo_url":"https://github.com/idt-iti/mmfusion-iml","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":1,"framework":"pytorch","reach":null}],"tasks":[{"task_slug":"detecting-image-manipulation","task_name":"Detecting Image Manipulation"},{"task_slug":"image-forensics","task_name":"Image Forensics"},{"task_slug":"image-forgery-detection","task_name":"Image Forgery Detection"},{"task_slug":"image-manipulation","task_name":"Image Manipulation"},{"task_slug":"image-manipulation-detection","task_name":"Image Manipulation Detection"},{"task_slug":"image-manipulation-localization","task_name":"Image Manipulation Localization"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[{"leaderboard":"/sota/image-manipulation-detection-on-coverage","task":"Image Manipulation Detection","dataset":"COVERAGE","model":"Early Fusion","rank_in_archive_order":1,"of":8,"metrics":{"AUC":".839","Balanced Accuracy":".770"},"uses_additional_data":false},{"leaderboard":"/sota/image-manipulation-detection-on-coverage","task":"Image Manipulation Detection","dataset":"COVERAGE","model":"Late Fusion","rank_in_archive_order":2,"of":8,"metrics":{"AUC":".792","Balanced Accuracy":".720"},"uses_additional_data":false},{"leaderboard":"/sota/image-manipulation-detection-on-casia-v1","task":"Image Manipulation Detection","dataset":"Casia V1+","model":"Late Fusion","rank_in_archive_order":1,"of":9,"metrics":{"AUC":".930","Balanced Accuracy":".860"},"uses_additional_data":false},{"leaderboard":"/sota/image-manipulation-detection-on-casia-v1","task":"Image Manipulation Detection","dataset":"Casia V1+","model":"Early Fusion","rank_in_archive_order":2,"of":9,"metrics":{"AUC":".929","Balanced Accuracy":".845"},"uses_additional_data":false},{"leaderboard":"/sota/image-manipulation-detection-on-cocoglide","task":"Image Manipulation Detection","dataset":"CocoGlide","model":"Late Fusion","rank_in_archive_order":1,"of":8,"metrics":{"AUC":".760","Balanced Accuracy":".677"},"uses_additional_data":false},{"leaderboard":"/sota/image-manipulation-detection-on-cocoglide","task":"Image Manipulation Detection","dataset":"CocoGlide","model":"Early Fusion","rank_in_archive_order":2,"of":8,"metrics":{"AUC":".755","Balanced Accuracy":".660"},"uses_additional_data":false},{"leaderboard":"/sota/image-manipulation-detection-on-columbia","task":"Image Manipulation Detection","dataset":"Columbia","model":"Early Fusion","rank_in_archive_order":2,"of":8,"metrics":{"AUC":".996","Balanced Accuracy":".962"},"uses_additional_data":false},{"leaderboard":"/sota/image-manipulation-detection-on-columbia","task":"Image Manipulation Detection","dataset":"Columbia","model":"Late Fusion","rank_in_archive_order":3,"of":8,"metrics":{"AUC":".977","Balanced Accuracy":".822"},"uses_additional_data":false},{"leaderboard":"/sota/image-manipulation-detection-on-dso-1","task":"Image Manipulation Detection","dataset":"DSO-1","model":"Early Fusion","rank_in_archive_order":1,"of":9,"metrics":{"AUC":".966","Balanced Accuracy":".935"},"uses_additional_data":false},{"leaderboard":"/sota/image-manipulation-detection-on-dso-1","task":"Image Manipulation Detection","dataset":"DSO-1","model":"Late Fusion","rank_in_archive_order":3,"of":9,"metrics":{"AUC":".958","Balanced Accuracy":".830"},"uses_additional_data":false},{"leaderboard":"/sota/image-manipulation-localization-on-coverage","task":"Image Manipulation Localization","dataset":"COVERAGE","model":"Early Fusion","rank_in_archive_order":1,"of":11,"metrics":{"Average Pixel F1(Fixed threshold)":".663"},"uses_additional_data":false},{"leaderboard":"/sota/image-manipulation-localization-on-coverage","task":"Image Manipulation Localization","dataset":"COVERAGE","model":"Late Fusion","rank_in_archive_order":2,"of":11,"metrics":{"Average Pixel F1(Fixed threshold)":".641"},"uses_additional_data":false},{"leaderboard":"/sota/image-manipulation-localization-on-casia-v1","task":"Image Manipulation Localization","dataset":"Casia V1+","model":"Early Fusion","rank_in_archive_order":2,"of":11,"metrics":{"Average Pixel F1(Fixed threshold)":".784"},"uses_additional_data":false},{"leaderboard":"/sota/image-manipulation-localization-on-casia-v1","task":"Image Manipulation Localization","dataset":"Casia V1+","model":"Late Fusion","rank_in_archive_order":3,"of":11,"metrics":{"Average Pixel F1(Fixed threshold)":".775"},"uses_additional_data":false},{"leaderboard":"/sota/image-manipulation-localization-on-cocoglide","task":"Image Manipulation Localization","dataset":"CocoGlide","model":"Late Fusion","rank_in_archive_order":2,"of":11,"metrics":{"Average Pixel F1(Fixed threshold)":".574"},"uses_additional_data":false},{"leaderboard":"/sota/image-manipulation-localization-on-cocoglide","task":"Image Manipulation Localization","dataset":"CocoGlide","model":"Early Fusion","rank_in_archive_order":4,"of":11,"metrics":{"Average Pixel F1(Fixed threshold)":".553"},"uses_additional_data":false},{"leaderboard":"/sota/image-manipulation-localization-on-columbia","task":"Image Manipulation Localization","dataset":"Columbia","model":"Early Fusion","rank_in_archive_order":1,"of":11,"metrics":{"Average Pixel F1(Fixed threshold)":".888"},"uses_additional_data":false},{"leaderboard":"/sota/image-manipulation-localization-on-columbia","task":"Image Manipulation Localization","dataset":"Columbia","model":"Late Fusion","rank_in_archive_order":4,"of":11,"metrics":{"Average Pixel F1(Fixed threshold)":".864"},"uses_additional_data":false},{"leaderboard":"/sota/image-manipulation-localization-on-dso-1","task":"Image Manipulation Localization","dataset":"DSO-1","model":"Late Fusion","rank_in_archive_order":2,"of":11,"metrics":{"Average Pixel F1(Fixed threshold)":".899"},"uses_additional_data":false},{"leaderboard":"/sota/image-manipulation-localization-on-dso-1","task":"Image Manipulation Localization","dataset":"DSO-1","model":"Early Fusion","rank_in_archive_order":4,"of":11,"metrics":{"Average Pixel F1(Fixed threshold)":".869"},"uses_additional_data":false}],"syntology":{"atlas_url":null,"mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}