Scholarlink

<p>The introduction of anti-image forensic operations puts a limit on the detection accuracy of different existing forensic detectors. It demands a robust forensic technique that can either expose forgery even if the anti-image forensic operation is applied or can detect the images that have gone through the anti-image forensic operation. This paper presents a machine learning approach for differentiating uncompressed images from different kinds of anti-forensically altered images. We propose a 576-dimensional feature for training and classification, which depends on the variation of the First Significant Digit (FSD) distribution of rounded-discrete cosine transform coefficients (R-DCT) from one subband to the next in zig-zag scanning order. The multi-class classification is done using a neural network based classifier. The quantitative experiments and analysis confirm that the proposed method achieves a good classification accuracy of 99.65%. The dimensionality of the proposed feature is further reduced with a slight fall in the accuracy of detection. The proposed approach can also be useful in the quality assessment of medical images and the validation of sensor data.</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

A Novel Image Forensics Approach Based on Machine Learning with use Case in Sensor Image Data Validation

  • 期刊论文

  • 发表日期：

    2025年8月6日

查看全文
  • Neeti Taneja,
  • Gouri Sankar Mishra,
  • Dinesh Bhardwaj

摘要

The introduction of anti-image forensic operations puts a limit on the detection accuracy of different existing forensic detectors. It demands a robust forensic technique that can either expose forgery even if the anti-image forensic operation is applied or can detect the images that have gone through the anti-image forensic operation. This paper presents a machine learning approach for differentiating uncompressed images from different kinds of anti-forensically altered images. We propose a 576-dimensional feature for training and classification, which depends on the variation of the First Significant Digit (FSD) distribution of rounded-discrete cosine transform coefficients (R-DCT) from one subband to the next in zig-zag scanning order. The multi-class classification is done using a neural network based classifier. The quantitative experiments and analysis confirm that the proposed method achieves a good classification accuracy of 99.65%. The dimensionality of the proposed feature is further reduced with a slight fall in the accuracy of detection. The proposed approach can also be useful in the quality assessment of medical images and the validation of sensor data.