Recompression can leave invisible traces that alter the integrity of digital evidence. Discover how our latest research is helping forensic analysts expose hidden tampering in HEVC video files.
We’re very glad to announce that Amped researchers have co-authored a scientific article about the detection of double video encoding for the modern HEVC codec. This paper, titled “Detection of Double Compression in HEVC Videos Containing B-Frames“ has been carried out together with the University of Florence (Italy). It was supervised by prof. Alessandro Piva and the outstanding contribution of Dr. Yoshihisa Furushita, as well as Dr. Daniele Baracchi and Dr. Dasara Shullani.
The article is freely accessible at this link.
Tackling the Complexity of HEVC Recompression
The paper addresses a scenario that is both common in real-world cases and underexplored in academic literature: how to reliably detect when a video has been re-encoded using HEVC with B-frames.
Why does this matter? In many forensic cases, investigators encounter video evidence that has undergone multiple compression stages, whether due to editing, re-uploading, or platform processing. Detecting such alterations is critical in establishing the integrity and chain of custody of the footage.
The Method: Learning from Compression Artifacts
We proposed a technique that models temporal inconsistencies introduced during double compression. Here’s how we approached the problem:
- Feature Extraction: From each frame, we extracted encoding features such as frame type, coding unit (CU) size, quantization parameter (QP), and prediction modes. These features were distilled into a 28-dimensional vector representing the entire video.
- Temporal Modeling: To analyze the sequence of frames over time, the study employed a Bidirectional Long Short-Term Memory (Bi-LSTM) neural network. This model is particularly well-suited for spotting subtle changes that emerge across frame sequences. These are exactly the kind of artifacts introduced during recompression.
Experimental Design and Results
To rigorously test their method, the researchers built a dedicated dataset of 129 HEVC-compressed YUV videos, derived from 43 original source sequences. This number was then largely expanded through data augmentation. Videos covered a range of bitrate combinations and GOP (Group of Pictures) structures to ensure variety and robustness.
The proposed approach achieved a detection accuracy of 80.06%, outperforming two established baselines. This makes it a promising tool for forensic analysts dealing with HEVC footage. This is especially true in surveillance and mobile device scenarios where B-frames are common and quality-preserving recompression is the norm.
What This Means for Forensics
At Amped Software, we understand the critical importance of preserving and interpreting the smallest traces left behind in digital evidence. This research reinforces the idea that encoding-level features can uncover hidden manipulations that are otherwise imperceptible.
As HEVC becomes increasingly widespread, so too must our ability to scrutinize it. Fortunately, forensic science is advancing with new innovations, and this research is pushing the field toward improved methods for detecting manipulated surveillance videos and verifying the integrity of footage.
We congratulate the authors from the University of Florence on their contribution and look forward to future advancements in this space.