Scholarlink

<p>Deepfake detection involves identifying artificial intelligence-generated synthetic media that manipulates audio, video, or images to distinguish authentic content from fake. It employs artificial intelligence algorithms, forensic analysis, and pattern recognition techniques to combat disinformation, fraud, and misuse. Existing techniques often struggle with dataset bias, limited generalization to unseen deepfakes, susceptibility to noise or compression artifacts, and high computational complexity. To overcome the problems above, this research introduces a Multimodal-weighted soft discernibility adversarial matrix Multi-Layer Perceptron model with olive ridley survival optimisation (MSMLP-ORS). The primary contributions of this research include enhancing model stability, ensuring result repeatability, and delivering consistent outcomes with high accuracy. The method begins with advanced preprocessing using Adaptive Kalman Non-Local Means denoising (AKNLM) to remove noise and optimally enhance the image quality. After the preprocessing, the features are extracted using three methods, namely gray level co-occurrence matrix, local binary pattern, and compact convolutional transformer. Then, the extracted images are fused and classified using MSMLP-ORS, and the optimization technique is used to improve the performance. To enhance interpretability and support ethical accountability, explainable AI is incorporated using Gradient-weighted Class Activation Mapping (Grad-CAM), which provides visual insights into the model’s predictions. Experiments are conducted with Python and achieved impressive outcomes in accuracy (99.7%), precision (99.9%), Recall (99.6%), <i>F</i>1-score (99.8%), and area under the curve (AUC) (1.0). The proposed model outperformed other models, such as refining localized attention features with multi-scale relationships network (AUC = 0.90), self-attention deep fake face discrimination network (AUC = 0.90), and multiscale vision transformer (AUC = 0.85) in all metrics.</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 with interpretability analysis using optimized algorithm

  • 期刊论文

  • 发表日期：

    2026年2月5日

查看全文
  • Jogi John,
  • Arvind M. Bhave,
  • Dhiraj Ghanshyam Karwatkar,
  • Rupali Prakash Gomase,
  • Sandhya D. Patil

摘要

Deepfake detection involves identifying artificial intelligence-generated synthetic media that manipulates audio, video, or images to distinguish authentic content from fake. It employs artificial intelligence algorithms, forensic analysis, and pattern recognition techniques to combat disinformation, fraud, and misuse. Existing techniques often struggle with dataset bias, limited generalization to unseen deepfakes, susceptibility to noise or compression artifacts, and high computational complexity. To overcome the problems above, this research introduces a Multimodal-weighted soft discernibility adversarial matrix Multi-Layer Perceptron model with olive ridley survival optimisation (MSMLP-ORS). The primary contributions of this research include enhancing model stability, ensuring result repeatability, and delivering consistent outcomes with high accuracy. The method begins with advanced preprocessing using Adaptive Kalman Non-Local Means denoising (AKNLM) to remove noise and optimally enhance the image quality. After the preprocessing, the features are extracted using three methods, namely gray level co-occurrence matrix, local binary pattern, and compact convolutional transformer. Then, the extracted images are fused and classified using MSMLP-ORS, and the optimization technique is used to improve the performance. To enhance interpretability and support ethical accountability, explainable AI is incorporated using Gradient-weighted Class Activation Mapping (Grad-CAM), which provides visual insights into the model’s predictions. Experiments are conducted with Python and achieved impressive outcomes in accuracy (99.7%), precision (99.9%), Recall (99.6%), F1-score (99.8%), and area under the curve (AUC) (1.0). The proposed model outperformed other models, such as refining localized attention features with multi-scale relationships network (AUC = 0.90), self-attention deep fake face discrimination network (AUC = 0.90), and multiscale vision transformer (AUC = 0.85) in all metrics.