Author(s): Salma Hamdy
Journal: International Journal of Computer Science and Information Security
ISSN 1947-5500
Volume: 9;
Issue: 9;
Start page: 36;
Date: 2011;
Original page
Keywords: Digital image forensics | forgery detection | compression history | Quantization tables.
ABSTRACT
A forensic analyst is often confronted with low quality digital images, in terms of resolution and/or compression, raising the need for forensic tools specifically applicable to detecting tampering in low quality images. In this paper we propose a method for quantization table estimation for JPEG compressed images, based on streamed DCT coefficients. Reconstructed dequantized DCT coefficients are used with their corresponding compressed values to estimate quantization steps. Rounding errors and truncations errors are excluded to eliminate the need for statistical modeling and minimize estimation errors, respectively. Furthermore, the estimated values are then used with distortion measures in verifying the authenticity of test images and exposing forged parts if any. The method shows high average estimation accuracy of around 93.64% against MLE and power spectrum methods. Detection performance resulted in an average false negative rate of 6.64% and 1.69% for two distortion measures, respectively.
Journal: International Journal of Computer Science and Information Security
ISSN 1947-5500
Volume: 9;
Issue: 9;
Start page: 36;
Date: 2011;
Original page
Keywords: Digital image forensics | forgery detection | compression history | Quantization tables.
ABSTRACT
A forensic analyst is often confronted with low quality digital images, in terms of resolution and/or compression, raising the need for forensic tools specifically applicable to detecting tampering in low quality images. In this paper we propose a method for quantization table estimation for JPEG compressed images, based on streamed DCT coefficients. Reconstructed dequantized DCT coefficients are used with their corresponding compressed values to estimate quantization steps. Rounding errors and truncations errors are excluded to eliminate the need for statistical modeling and minimize estimation errors, respectively. Furthermore, the estimated values are then used with distortion measures in verifying the authenticity of test images and exposing forged parts if any. The method shows high average estimation accuracy of around 93.64% against MLE and power spectrum methods. Detection performance resulted in an average false negative rate of 6.64% and 1.69% for two distortion measures, respectively.