Scholarlink

<p>The rapid advancement of generative models has precipitated a surge in synthetic imagery, posing a severe threat to the integrity of digital information across multiple domains. Detecting such content presents a significant challenge, particularly in low-resolution scenarios where manipulation artifacts are often subtle or obscured. To address this issue, we propose a comprehensive Ensemble Learning framework designed to enhance deepfake detection capabilities on the CIFAKE dataset. We conduct a systematic empirical evaluation of four core strategies: bagging, voting, boosting, and stacking, utilizing state-of-the-art CNN architectures as backbone models. Furthermore, this study introduces a dynamic threshold optimization (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\tau ^*\)</EquationSource> </InlineEquation>) and weight tuning technique to minimize false positive rates. Experimental results demonstrate that the Bagging strategy outperforms existing benchmarks. Concurrently, our analysis confirms that leveraging architectural diversity is more effective than relying on complex single models, offering an optimal balance between detection accuracy and computational efficiency.</p>

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

Deepfake detection on low-resolution images using optimized Ensemble Learning

  • 期刊论文

  • 发表日期：

    2026年8月4日

查看全文
  • Linh Thuy Thi Pham,
  • Cu Vinh Loc,
  • Dang Duy Huynh,
  • Hai Thanh Nguyen

摘要

The rapid advancement of generative models has precipitated a surge in synthetic imagery, posing a severe threat to the integrity of digital information across multiple domains. Detecting such content presents a significant challenge, particularly in low-resolution scenarios where manipulation artifacts are often subtle or obscured. To address this issue, we propose a comprehensive Ensemble Learning framework designed to enhance deepfake detection capabilities on the CIFAKE dataset. We conduct a systematic empirical evaluation of four core strategies: bagging, voting, boosting, and stacking, utilizing state-of-the-art CNN architectures as backbone models. Furthermore, this study introduces a dynamic threshold optimization ( \(\tau ^*\) ) and weight tuning technique to minimize false positive rates. Experimental results demonstrate that the Bagging strategy outperforms existing benchmarks. Concurrently, our analysis confirms that leveraging architectural diversity is more effective than relying on complex single models, offering an optimal balance between detection accuracy and computational efficiency.