Scholarlink

<p>Deepfakes pose growing risks to cybersecurity, social engineering, and media integrity, yet existing public datasets provide limited coverage of underrepresented English accents and synchronised audio–visual content. To address this gap, we present <i>IrishAV-Deepfake</i>, two Irish-accented English deepfake datasets designed for multimodal deepfake detection research. The collection includes synchronised audio–video and corresponding audio-only datasets containing both authentic and synthetically generated samples. Authentic Irish-accented media were retrieved from publicly available sources using Irish-related metadata and manually verified through auditory inspection and speaker/source context review. Synthetic audio samples were generated using Google Text-to-Speech and Hugging Face Parler-TTS with Irish-accent-oriented settings, while synthetic video samples were created using a prompt-based text-to-video workflow. All media files underwent DPI-inspired file-level integrity validation, including file signature verification, format checks, size constraints, and metadata consistency checks. The datasets contain balanced real and fake partitions with gender annotations to support robustness, bias, and generalisation studies. All data, code, and documentation are publicly released to support reproducibility and downstream experimentation.</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

Irish-Accented English Audio-Visual Deepfake Datasets with Deep Packet Inspection-Inspired Media Integrity Validation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月30日

查看全文
  • Teerath Kumar,
  • Muhammad Umair,
  • Susan Leavy,
  • Patrick Eustace,
  • Edward Curry,
  • Mamoona Naveed Asghar

摘要

Deepfakes pose growing risks to cybersecurity, social engineering, and media integrity, yet existing public datasets provide limited coverage of underrepresented English accents and synchronised audio–visual content. To address this gap, we present IrishAV-Deepfake, two Irish-accented English deepfake datasets designed for multimodal deepfake detection research. The collection includes synchronised audio–video and corresponding audio-only datasets containing both authentic and synthetically generated samples. Authentic Irish-accented media were retrieved from publicly available sources using Irish-related metadata and manually verified through auditory inspection and speaker/source context review. Synthetic audio samples were generated using Google Text-to-Speech and Hugging Face Parler-TTS with Irish-accent-oriented settings, while synthetic video samples were created using a prompt-based text-to-video workflow. All media files underwent DPI-inspired file-level integrity validation, including file signature verification, format checks, size constraints, and metadata consistency checks. The datasets contain balanced real and fake partitions with gender annotations to support robustness, bias, and generalisation studies. All data, code, and documentation are publicly released to support reproducibility and downstream experimentation.