Abstract: Manipulated videos are frequently an important part of misinformation campaigns. While much attention has recently focused on sophisticated threats such as deepfakes, the majority of videos used in misinformation campaigns have been created using relatively simple manipulations that can be produced by commonly available editing software. One important manipulation that has been used in previous misinformation attempts is altering the speed of a video. In the previous 18 months, widely circulated videos have had their speed manipulated to make Speaker of the House Nancy Pelosi appear disoriented, and to make reporter Jim Acosta appear to act aggressively toward a White House staffer. Currently, however, there are no approaches to accurately detect video speed manipulation that can be deployed at scale. In this paper, we propose new algorithms to detect video speed manipulation and to estimate the rate by which a video's speed has been modified. To do this, we identify a new trace left by video speed manipulation and show how it can be extracted from a video. Our approaches to trace extraction, speed manipulation detection, and manipulation rate estimation are computationally efficient and can be run in a matter of milliseconds. We present experimental results that show that our proposed approach can detect manipulated videos with up to 99% accuracy.
The Rise of "Cheapfakes"
While the public remains fixated on the hyper-realistic threat of AI-generated deepfakes, a simpler form of deception is already impacting public discourse. These "cheapfakes"—videos sped up or slowed down to alter a subject's perceived demeanor—have been used to make public figures appear intoxicated or aggressive. Notable examples include a slowed-down video of Nancy Pelosi intended to make her appear disoriented and a sped-up clip of reporter Jim Acosta intended to make him appear aggressive.
Traditional forensic tools are often too slow for social media scale or focus on pixel-level artifacts that are easily masked. Researchers Brian C. Hosler and Matthew C. Stamm from Drexel University have developed a lightweight detection method that flags these manipulations across millions of uploads in real-time by analyzing the video bitstream metadata rather than the pixels themselves.
Identifying the Forensic Trace
The core insight of this research is that video editing software rarely performs expensive frame interpolation when changing speed. Instead, it typically duplicates frames to slow down or deletes frames to speed up. When these edited videos are re-encoded (e.g., using the H.264/AVC standard), these duplicates and deletions leave a distinct "fingerprint" in the Encoded Frame Size (EFS)—the number of bytes used to store each frame.
In a slowed video, repeated frames result in near-zero residuals because the frame is identical to the last, creating periodic "drops." In a sped-up video, skipped frames result in periodic spikes because the "jump" in motion makes the next frame harder to predict and compress.
Methodology: From Bitstream to Classification
The proposed system extracts the EFS sequence directly from the video bitstream without fully decoding the video, making it exceptionally fast. The process involves three main stages:
- Signal Conditioning: Because frame sizes vary naturally based on frame type (I, P, or B frames), the EFS is split into three subsequences. Median filtering is used to estimate the "ideal" frame size, which is then subtracted to isolate a residual signal.
- Modeling: An Nth-order Auto-Regressive (AR) model captures the periodicities within these residuals.
- Classification: AR coefficients are fed into a Support Vector Machine (SVM) for binary classification (manipulated vs. original) and a Support Vector Regressor (SVR) to estimate the specific speed manipulation parameter (SMP).
Key Results and Real-World Application
The system achieved a maximum accuracy of 99.1% for detecting slowed-down videos. When tested on "organic" videos—those edited with consumer-grade software rather than command-line tools—the accuracy remained high at 99.3%. Notably, in a case study of the infamous Pelosi video, the model successfully identified the manipulation with 100% accuracy across all tested segments.
Beyond accuracy, the method is incredibly efficient. Feature extraction and classification take only about 0.001 seconds per video, making it viable for massive-scale deployment on social media backends. While detection performance dips slightly as the speed change approaches 1.0x (where the trace is weakest), the approach provides a critical first line of defense against low-effort, high-impact misinformation.
Summary updated