Mathematical Analysis of Backdoor-Induced Melanoma Risk in Modality-Heterogeneous Federated Learning
Open Access
- 23 October 2026
- journal article
- research article
- Published by MDPI AG in Mathematics
- Vol. 14 (21) , 3870
Abstract
Aggregate accuracy and conventional attack success rates cannot distinguish pre-existing diagnostic errors from melanoma false negatives induced by a trigger. We evaluate melanoma-to-nevus risk downgrading in source-heterogeneous federated learning using paired clean and triggered predictions. Four source-defined clients represent clinical photography, smartphone imaging, and dermoscopy. The primary endpoint, trigger-induced paired false negatives (trigFN), counts clean-correct melanoma cases that change to a nevus prediction after triggering. In centralized screening, sticker and blend triggers produce 35 and 31 trigFN among 77 melanoma cases, whereas a square trigger produces none. Across four seeds, Full-Phase poisoning produces trigFN and a -percentage-point melanoma-sensitivity loss; Late-Phase poisoning produces trigFN and a -point loss. Under Full-Phase poisoning, NormClip with reduces trigFN to and the sensitivity loss to points, while mean balanced accuracy (BACC) changes from to . External evaluation shows smaller within-model trigger-associated losses with clipping, although absolute melanoma sensitivity on untriggered images is lower. Paired clinical endpoints reveal harm that aggregate utility can obscure. Within the evaluated four-client, single-attacker, non-adaptive setting, norm-constrained aggregation reduces but does not eliminate harm.
Keywords
Funding Information
- Changzhou Science and Technology Program (CJ20259010)
This publication has 15 references indexed in Scilit:
- Predicting anemia risk in Iranian adults without diagnostic blood indices: an interpretable machine learning approach integrating dietary, clinical, and non-diagnostic laboratory patternsScientific Reports, 2026
- A causal machine learning framework for ICU discharge decision support using electronic health recordsScientific Reports, 2026
- FIBA: Frequency-Injection based Backdoor Attack in Medical Image Analysis2022
- Federated learning for predicting clinical outcomes in patients with COVID-19Nature Medicine, 2021
- Evaluating Deep Neural Networks Trained on Clinical Images in Dermatology with the Fitzpatrick 17k Dataset2021
- PAD-UFES-20: A skin lesion dataset composed of patient data and clinical images collected from smartphonesData in Brief, 2020
- The future of digital health with federated learningnpj Digital Medicine, 2020
- Federated learning in medicine: facilitating multi-institutional collaborations without sharing patient dataScientific Reports, 2020
- The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesionsScientific Data, 2018
- Seven-Point Checklist and Skin Lesion Classification Using Multitask Multimodal Neural NetsIEEE Journal of Biomedical and Health Informatics, 2018