Detection of frame deletion for digital video forensics

TLDR: This paper examines the authenticity of digital video evidence and in particular it proposes a machine learning approach to detecting frame deletion and it is shown that the proposed solution works for detecting forged videos regardless of the number of deleted frames.

About: This article is published in Digital Investigation.The article was published on 2013-12-01 and is currently open access. It has received 72 citations till now. The article focuses on the topics: Video quality & Video compression picture types.

Figures

Figure 1. Block diagram of proposed frame deletion detection

Table 2. Features retained after applying stepwise regression for P and B frames

Table 3. True positive and false alarm rates (%) for frame deletion using a fixed GoP length.

Table 1. List of features computed from P and B frames

Figure 3. Histogram of projected features using spectral regression.

Table 7. Average and standard deviation of classification rates of the proposed solution, CBR coding used.

Frequently asked questions

Q1. What are the contributions mentioned in the paper "Detection of frame deletion for digital video forensics" ?

This paper examines the authenticity of digital video evidence and in particular it proposes a machine learning approach to detecting frame deletion. Consequently, the dimensionality of the feature vectors is reduced using spectral regression where it is shown that the projected features of unaltered and forged videos are nearly separable. 

Q2. What future works have the authors mentioned in the paper "Detection of frame deletion for digital video forensics" ?

In future work, the proposed solution can be modified and extended in order to determine the exact location of the deleted frames and not just detect the existence of frame deletion. 

Q3. What is the use of z-scores for normalizing a video?

The mean and standard deviation vectors of the training feature vector set are stored and used for normalizing a feature vector for a video under examination. 

Q4. Why do courts use forensics experts to assess the quality and authenticity of digital video evidence?

Because of inherent visual artifacts and possible fraud, courts usually call upon the testimony of forensics experts to subjectively assess the quality and authenticity of the digital video evidence. 

Q5. How can a video be detected by examining the distribution of quantized DCT coefficients?

Double compression of MPEG-2 video can also be detected by examining the distribution of quantized DCT coefficients as proposed in [7]. 

Q6. What is the QCIF codec used to compress?

An MPEG-2 codec which is an implementation of the ISO/IEC DIS 13818-2 international standard [18] is used to compress 36 standard QCIF test sequences using VBR and CBR coding. 

Q7. What is the importance of examining the authenticity of the compressed video?

Prior to the admissibility of the compressed video to a courtroom it is of prime importance to examine the authenticity of the compressed video. 

Q8. What is the way to assess the quality of a video?

Assessing the quality of the compressed video in the absence of the reference can be performed on the whole video as one unit thus quantifying its quality [1]. 

Q9. What is the way to measure the false alarm rates?

In the following experiments, the true positive (TP) rates and the false alarm rates (FA) are reported using KNN (with K set to 1 with an Euclidean distance similarity measure), logistic regression and SVM with a quadratic kernel. 

Q10. What is the mean prediction residual energy of MBs in a video sequence?

The mean prediction residual energy of non-Intra coded MBs is computed using Equation 4:𝜇𝐸 = 1𝑁⁄ ∑ ∑ 𝑃𝑖(𝑗),𝑖𝑗 𝑖 ∈ {𝑖𝑛𝑑𝑖𝑐𝑒𝑠 𝑜𝑓 𝑛𝑜𝑛 𝑖𝑛𝑡𝑟𝑎 𝑀𝐵𝑠 𝑎𝑡 𝑡ℎ𝑒 𝑗𝑡ℎ 𝑓𝑟𝑎𝑚𝑒} (4)Where N is the total number of predicted MBs in the video sequence for a P or a B frame. 

Q11. What is the way to determine the T value of the alternative hypothesis?

If the corresponding T value becomes insignificant (i.e. the alternative hypothesis 𝐻1 is rejected.), 𝑥1 is removed and the predictors are searched for a variable that generates the highest T value in the presence of 𝛽2𝑥2. 

Q12. What is the procedure for removing predictors from the model?