Elsevier

Neurocomputing

Volume 514, 1 December 2022, Pages 231-244
Neurocomputing

Detecting double H.266/VVC compression with the same coding parameters

https://doi.org/10.1016/j.neucom.2022.09.153Get rights and content

Abstract

The detection of video double compression with the same coding parameters is a difficult problem in video forensics, since the recompression traces are extremely slight in this case. Although a series of methods have been proposed, detection of Versatile Video Coding (H.266/VVC) double compression with the same coding parameters is rarely reported. To address this issue, we propose a novel algorithm for the detection of H.266/VVC double compression with the same coding parameters in this paper. We first analyze the generation of H.266/VVC double compression traces, and explore the variation of coding modes caused by multiple compressions. Then we define Minimum Unit Mapping (MUM) and Subunit Prediction Mapping (SPM) to facilitate feature construction. The strength of the variation in the number of coding unit (CU) in each compression is calculated to form CU partition modes based feature set. We also use the consistency ratios of adjacent prediction mode pairs in the horizontal, vertical, major diagonal, and minor diagonal directions after multiple compressions to obtain prediction modes based feature set. These feature sets are further concatenated into a fusion feature for detection. With the thorough comparison to the state-of-the-art algorithms, it is verified that the proposed algorithm outperforms other current works in the field. Besides, the proposed method is robust against various encoding configurations.

Introduction

The emergence of video editing tools has brought great convenience to our daily life, however, it also has a negative impact on the authenticity and originality of digital video [1], [2], [3], [4]. Therefore, it is worthwhile to apply passive video forensics method to detect video manipulation and editing.

As a kind of video editing, video double compression can recompress a video into any specific quality. The detection of video double compression has received more and more attention in recent years [5], [6]. For example, in 2009, Wang et al. [7] pointed out that the distribution of quantized transformation coefficients would change if a video undergoes double compression. They then utilized this feature for media forensics. Similarly, many scholars tried to explore the statistical features in coding domain to detect double compressed video. In [8], Chen et al. introduced the first digit distribution of non-zero quantized Alternating Current (AC) coefficients for the detection. Jiang et al. [9] utilized the Markov properties of transformation coefficients in four different directions to construct fusion feature. Promising performance is shown in the case of different compression qualities. More works [10], [11], [12] using the distributions of Discrete Cosine Transform (DCT) coefficients have also been proposed to tackle the issue. Since video double compression can also lead to the variation of the macroblock prediction types, Vázquez-Padín et al. [13] proposed an algorithm based on the variation in inter frames to detect double compressed Moving Picture Experts Group (MPEG) videos. These authors provided a deep theoretical analysis allowing us to understand the artifacts that justify the presence of the Variation of Prediction Footprint (VPF) in double compressed videos in [14] and put forward a generalized variation of prediction footprint for double compression detection. Since the prediction strategy of High Efficiency Video Coding (H.265/HEVC) differs from its predecessor to some extent, Xu et al. proposed the Sequence of Number of Prediction Unit of its Prediction Mode (SN-PUPM) feature for the detection of H.265/HEVC double compression [15], and achieves an Area Under Curve (AUC) of 0.9226. The inherent blocking artifacts in video coding also motivate scholars to explore traces in these artifacts for detection. In [4], He et al. proposed an adaptive post-filtering technique to measure the strength of block artifacts in the decompression domain. The strength sequences were used to detect double compression and estimate the size of the original Group of Pictures (GOP). Xu et al. [16] pointed out that the filtering mode in single compressed HEVC video is different from that of the transcoded video. They then adopted in-loop filtering to detect double compressed H.265/HEVC videos. Results proved that the algorithm outperforms the compared algorithms and has better robustness. The same authors also evaluated the impact of video motion information on double compression detection, and proposed a motion-adaptive detection algorithm in [17]. In addition, some deep learning-based double compression detection methods have also been proposed in recent years. For example, in [18], He et al. proposed a detection method for fake bitrate videos using a hybrid deep-learning network from recompression error. The authors designed a network that contains two branches with heterogeneous structures for the detection, the average detection accuracy for different bitrate combinations is above 92%. Besides, genetic Convolutional Neural Networks (CNN) is introduced in [19] to detect relocated I-frames. More statistical features like prediction residual [20], [21], prediction mode [22], [23], [24], and the idempotency property [25], [26] are also adopted to detect double compressed videos.

Although these algorithms work well when the parameters of double compression are inconsistent, some scholars found that when the video undergoes double compression with the same coding parameters, the detection would be more difficult. The reason is that when the same encoding parameters are used for video recompression, the difference between the coded videos is weaker than that of the recompression with different parameters, and the recompression traces are much more inconspicuous in this case. To tackle this problem, many scholars have proposed related algorithms in recent years. In [27], Chen et al. defined the concept of MacroBlock Mode (MBM) feature for the detection of MPEG double compression with the same Quantization Parameter (QP). Jiang et al. [28] made some improvements on this basis by calculating the variation rate of the unstable block, so that the algorithm can be applied to Advanced Video Coding (AVC) video. The general procedure includes multiple compression and decompression, intensities calculation, and classification. Results proved the robustness of these algorithms against various compression qualities. However, they cannot be directly applied to H.265/HEVC videos due to the upgrade of the block partitioning modes. Then, in [29], Intra Prediction Unit Prediction Mode (IPUPM) is adopted to evaluate the block stability. By calculating the variation rates of PUs, experimental results demonstrate the effectiveness of the algorithm for H.265/HEVC double compression detection.

In summary, the above mentioned works can achieve promising results under previous coding standards. However, when facing the latest Versatile Video Coding (H.266/VVC) standard, there will be a high probability of performance degradation or execution failure. It is because the most recent international video coding standard, H.266/VVC, was developed to serve an ever-growing need for improved video compression as well as to support a wider variety of today’s media content and emerging applications [30]. The new coding standard has introduced a series of new coding techniques, such as multi-type tree partitioning, Adaptive Loop Filtering (ALF) and so on, which inevitably interferes with the feature extraction of existing algorithms. However, no relevant algorithm for H.266/VVC double compression detection has been proposed so far. Therefore, it is of great significance to develop a novel algorithm for the detection of H.266/VVC double compression.

In this paper, we propose an algorithm to detect H.266/VVC double compression. Specifically, our work offers the following contributions: (1) We study the new problem of video double compression detection for H.266/VVC standard by leveraging modes analysis in the coding domain. (2) We calculate the variation of partition modes and consistency of prediction modes of coding units after consecutive compressions and decompressions to characterize the double compression traces. The variation in the number of CUs with different partition modes in each compression and the consistency ratios of adjacent prediction mode pairs in four directions are calculated to form the detection feature. (3) Minimum Unit Mapping (MUM) and Subunit Prediction Mapping (SPM) are designed to facilitate the feature construction. We use MUM to reduce the interference of quadtree or multi-type tree partition to mode analysis. SPM is adopted so that the number of adjacent prediction mode pairs in different directions can be calculated more accurately. (4) Extensive experiments are performed. A wide variety of scenarios and parameter settings are introduced to evaluate the effectiveness and robustness of the proposed algorithm comprehensively.

The remainder of this paper is organized as follows: In Section 2, new features of the H.266/VVC standard and in-depth analysis of the generation of H.266/VVC double compression traces are provided, while in Section 3 we introduce the detection algorithm. The experimental validation and comparison with existing methods are provided in Section 4. Finally, the conclusions are drawn in Section 5.

Section snippets

Quadtree and multi-type tree structures in H.266/VVC

H.266/VVC still adopts block-based hybrid coding strategy like its predecessor H.265/HEVC, and the basic processing unit within a picture in H.266/VVC is the coding tree unit (CTU) [30]. For the block partition strategy, a picture can be divided into Coding Tree Units (CTUs), CTUs can be further divided into several coding units (CUs) by quadtree or multi-type tree partitioning with a size range: the Smallest CU (SCU) size of 8×8 to the Largest CU (LCU) size of 128×128 [31]. The definition of

CU partition in multiple compressed video

Video multiple compression, including H.266/VVC double compression, would inevitably introduce additional traces. These traces would definitely affect the rate-distortion optimization results and have an impact on the encoded output to some extent. In this section, we analyze the variation of CU partition after H.266/VVC multiple compression to verify this point of view and to promote the exploration of reasonable feature characterization, so as to design effective detection features and

Experiments and analysis

After introducing the main steps of the proposed algorithm, in this section, extensive experiments are conducted for fair comparison to other work on H.266/VVC double compression detection. Regarding the robustness against more realistic scenarios, such as Constant Quantization Parameter (CQP), Average Bit Rate (ABR) rate control modes, different GOP sizes and spatial resolutions are verified.

Conclusion

In this paper, a novel algorithm is proposed for the detection of H.266/VVC double compression with the same coding parameters. Based on the in-depth analysis of the generation of H.266/VVC double compression traces, we find that the quality degradation characteristic still exists in H.266/VVC double compression. Then, the variation of CU partition and consistency of prediction modes are exploited to characterize double compression traces. In order to reduce the interference of quadtree or

CRediT authorship contribution statement

Qiang Xu: Conceptualization, Methodology, Writing - original draft. Dongmei Xu: Writing - review & editing, Validation, Visualization. Hao Wang: Software. Zhongjie Mi: Data curation. Zhe Wang: Investigation. Hong Yan: Supervision.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgements

This work is supported by Hong Kong Innovation and Technology Commission (InnoHK Project CIMDA), Hong Kong Research Grants Council (Project 11204821), City University of Hong Kong (Project 9610034), and Chongqing Natural Science Foundation (Project cstc2020jcyj-msxmX0635).

Qiang Xu received the Ph.D. degree in Cybersecurity from Shanghai Jiao Tong University, Shanghai, China, in 2021. He is currently a postdoctoral research fellow at City University of Hong Kong. His research interest includes multimedia forensics and security.

References (38)

  • W. Chen et al.

    Detection of double MPEG compression based on first digit statistics

  • X.H. Jiang et al.

    Detection of double compression in MPEG-4 videos based on markov statistics

    IEEE Signal Processing Letters

    (2013)
  • Y.T. Su et al.

    Detection of double-compression in MPEG-2 videos

  • J.Y. Xu et al.

    Detection of double MPEG-2 compression based on distributions of DCT coefficients

    International Journal of Pattern Recognition and Artificial Intelligence

    (2013)
  • M.L. Huang et al.

    Detection of double compression for HEVC videos based on the co-occurrence matrix of DCT coefficients

  • D. Vazquez-Padin et al.

    Detection of video double encoding with gop size estimation

  • D. Vázquez-Padín et al.

    Video integrity verification and gop size estimation via generalized variation of prediction footprint

    IEEE Transactions on Information Forensics and Security

    (2019)
  • Q. Xu et al.

    Hevc double compression detection based on sn-pupm feature

  • Q. Xu et al.

    Motion-adaptive detection of hevc double compression with the same coding parameters

    IEEE Transactions on Information Forensics and Security

    (2022)
  • Cited by (0)

    Qiang Xu received the Ph.D. degree in Cybersecurity from Shanghai Jiao Tong University, Shanghai, China, in 2021. He is currently a postdoctoral research fellow at City University of Hong Kong. His research interest includes multimedia forensics and security.

    Dongmei Xu works at the Department of Ophthalmology, Xingguo People’s Hospital, Jiangxi, China. Her research interests include medical image processing, multimedia content analysis.

    Hao Wang is currently an Associate Professor with the College of Computer Science and Technology, Chongqing University of Posts and Telecommunications, Chongqing, China. His current research interests include multimedia forensics, information security and privacy preserving.

    Zhongjie Mi is currently pursuing the Ph.D. degree with the School of Electronic Information and Electrical Engineering, Shanghai Jiao Tong University, Shanghai, China. His research interests include facial recognition, multimedia forensics and identity security.

    Zhe Wang received the Ph.D. degree from the Department of Electrical Engineering, Tsinghua University, Beijing, China, in 2021. He is currently a postdoctoral research fellow at the Department of Advanced Design and Systems Engineering, City University of Hong Kong, Hong Kong SAR. He is also with the Center for Intelligent Multidimensional Data Analysis, Hong Kong Science Park, Shatin, Hong Kong SAR. His current research interests include Structural Health Monitoring, Deep Learning, and Prognostics and Health Management.

    Hong Yan is currently a chair professor of computer engineering and Wong Chung Hong professor of data engineering at the City University of Hong Kong, Hong Kong. His research interests include image processing, pattern recognition and bioinformatics. Professor Yan is an IEEE Fellow and IAPR Fellow, and he received the 2016 Norbert Wiener Award from the IEEE Systems, Man and Cybernetics Society for contributions to image and biomolecular pattern recognition techniques.

    View full text