Scholarlink

Automatic detection of diabetic retinopathy (DR) is crucial, as this technology can significantly reduce diagnostic time. Given the privacy protection characteristics of federated learning (FL), a federated learning DR detection technology that can effectively protect patient privacy is very popular. However, attackers can still exploit privacy information through gradient attacks, data poisoning, and other means. To address privacy concerns, we introduce a novel DR detection technique called FedADP. In FedADP, we integrate differential privacy (DP) into our self-designed federated learning adaptive threshold optimization method. We provide a complete set of hospital collaboration processes trained using FL. Furthermore, this paper presents a Flexible Model Calibration Strategy (FMCS) to mitigate model drift issues caused by data heterogeneity. The experimental results validate the superiority of our method over other FL methods in terms of privacy protection. Compared to the federated averaging algorithm (FedAvg), FMCS effectively resolves the challenges arising from non-independent and identically distributed data. Moreover, compared to other baselines, FedADP provides stronger guarantees of privacy protection and achieves superior detection results.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Differential Privacy Federated Learning Approach for Diabetic Retinopathy Detection

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Yicheng Li,
  • Yijia Zhang,
  • Guantong Liu,
  • Mingyu Lu

摘要

Automatic detection of diabetic retinopathy (DR) is crucial, as this technology can significantly reduce diagnostic time. Given the privacy protection characteristics of federated learning (FL), a federated learning DR detection technology that can effectively protect patient privacy is very popular. However, attackers can still exploit privacy information through gradient attacks, data poisoning, and other means. To address privacy concerns, we introduce a novel DR detection technique called FedADP. In FedADP, we integrate differential privacy (DP) into our self-designed federated learning adaptive threshold optimization method. We provide a complete set of hospital collaboration processes trained using FL. Furthermore, this paper presents a Flexible Model Calibration Strategy (FMCS) to mitigate model drift issues caused by data heterogeneity. The experimental results validate the superiority of our method over other FL methods in terms of privacy protection. Compared to the federated averaging algorithm (FedAvg), FMCS effectively resolves the challenges arising from non-independent and identically distributed data. Moreover, compared to other baselines, FedADP provides stronger guarantees of privacy protection and achieves superior detection results.