Moiré Video Authentication: A Physical Signature Against AI Video Generation¶

Conference: ECCV 2026
arXiv: 2604.01654
Project Page: https://yuanqing-ai.github.io/physical_video_signature/
Code: None
Area: AI Security
Keywords: Moiré patterns, Video authentication, AI-generated video detection, Physical signature, Optical invariants

TL;DR¶

This paper proposes using moiré interference patterns as a physical signature to authenticate video authenticity: a compact double-layer grating structure is placed in the shooting scene, and the Pearson correlation coefficient between the fringe phase change and the translational displacement of the grating image is calculated to distinguish real videos from AI-generated ones. In real videos, these two are strictly coupled by the laws of optical geometry (mean correlation coefficient of 0.87), whereas even the strongest current video generation models (Veo 3.1, Grok Imagine, LTX-2) fail to precisely reproduce this coupling even under optimal configurations (mean correlation coefficient of 0.57, Cohen's d of 1.71, indicating highly significant differences).

Background & Motivation¶

AI-generated video is penetrating the information ecosystem at an alarming rate. From Sora and Veo to Grok Imagine and LTX-2, the videos produced by these models have become so realistic that they are difficult to distinguish for the human eye, automated detectors, and even forensic analysts. In 2024, a multinational company employee in Hong Kong was defrauded of $25 million during a video conference by deepfaked executive faces; in 2025, an Irish presidential candidate was impersonated by an AI-generated withdrawal statement video that spread on social media for hours—these are not alarmist scenarios but actual occurrences. Existing video authentication methods fall into a dilemma. Passive detectors based on post-hoc forensics (e.g., FaceForensics, CNN-based fingerprinting) are caught in an arms race with generative models: each new generation of models renders old detectors obsolete by eliminating the artifacts they rely on, necessitating continuous retraining and failing to provide long-term reliability guarantees. Digital watermarking schemes (e.g., C2PA metadata, SynthID), although able to embed provenance information at the time of creation, suffer from metadata stripping or forgery, and watermarks degrade under common post-processing operations.

The common flaw of both categories of methods is their failure to establish an unforgeable physical link from the captured scene to the final video. This paper proposes a fundamental paradigm shift: instead of attempting to "detect" artifacts in generated videos, it lets a real camera naturally produce a physical signal that generative models cannot precisely replicate. The key insight is that real videos are generated by optical systems adhering to physical laws, while generative models learn statistical correlations from data, rarely modeling the underlying physical processes. If a physical phenomenon exists whose appearance is tightly coupled with exact optical geometry, it can serve as a signature naturally generated by real cameras but synthesized incorrectly by current generative models. This work chooses moiré patterns: the interference patterns produced when two fine periodic structures (line gratings) overlap are extremely sensitive to the geometric relationship between the camera and the gratings. Minor camera movements trigger deterministic displacements of the fringes, which follow strict optical geometric formulas. Core Idea: Using the moiré motion invariant as an authentication criterion—fringe phase changes and grating image translational displacements are linearly coupled by optical geometry, with a theoretical correlation coefficient close to 1, regardless of the viewing distance and grating structural parameters. This coupling naturally holds in real videos, but in AI-generated videos, because generative models do not solve optical equations, they fail to reproduce this linear relationship, resulting in significantly lower correlation coefficients, thereby enabling video authenticity verification without any prior calibration.

Method¶

The proposed authentication system consists of a compact three-layer grating component (front lenticular film + middle printed fringes + bottom acrylic substrate) and an optically derived three-stage video analysis pipeline. The user wears or places the grating component in the shooting scene (e.g., a speaker's badge or a background sign at a press conference), where natural human or camera movement generates relative displacement. The pipeline independently extracts the cumulative phase signal of the fringes and the pure translational displacement signal of the grating component from the video, calculating their correlation coefficient as the authentication criterion.

Overall Architecture¶

The authentication pipeline takes a video containing the grating component as input and outputs an authentication decision after three core processing stages. The output of the first stage is simultaneously fed into the second and third stages, where the two signals ultimately converge.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Input Video"] --> B["Fringe Detection & Rectification<br/>ArUco Localization → Perspective Correction → CLAHE Enhancement → 1D Profile"]
    B --> C["Phase Tracking<br/>FFT Peak Frequency Locking → Phase Extraction → Incremental Unwrapping → Cumulative Phase Φ(t)"]
    B --> D["Motion Decomposition<br/>PnP Pose Estimation → Separation of Rotation Component → Translational Displacement Δu<sub>trans</sub>"]
    C --> E["Pearson Correlation Coefficient<br/>30-frame Sliding Window"]
    D --> E
    E --> F["|ρ| ≥ Threshold → Real Video<br/>|ρ| < Threshold → Generated Video"]

Phase 1 (Fringe Detection & Rectification): Detects and tracks the ArUco markers at the four corners of the grating component to compute a homography matrix, rectifying the fringe region to a frontal canonical view. After CLAHE contrast enhancement, it is averaged along the fringe direction to yield a one-dimensional intensity profile. Phase 2 (Phase Tracking): Performs a 1D FFT on the intensity profile to lock onto the dominant frequency component of the moiré beat frequency. It extracts the phase angle frame-by-frame and accumulates it into a smooth cumulative phase signal $\Phi(t)$ via incremental phase unwrapping. Phase 3 (Motion Decomposition & Authentication): Performs PnP pose estimation using the 3D-2D correspondences provided by the ArUco markers to separate the rotational components from the raw center trajectory (since rotation does not yield parallax and thus does not affect the fringe phase), retaining only the translational displacement $\Delta u_{\text{trans}}$ projected onto the direction sensitive to the fringes. Finally, the Pearson correlation coefficient between $\Phi(t)$ and $\Delta u_{\text{trans}}$ is calculated over a 30-frame sliding window.

Key Designs¶

1. Moiré Motion Invariant: The Theoretical Cornerstone of Distance Independence

The core theoretical challenge of moiré-based video authentication is that the distance $D$ between the camera and the grating is unknown and time-varying, and the fringe phase change depends on both the camera displacement and the distance. Directly comparing the phase change would require prior distance calibration. The key theoretical breakthrough of this work is the discovery of an invariant independent of $D$. For a front grating period $p_f$, a rear grating period $p_r$, and an interlayer gap $g$, a lateral camera displacement $\Delta x_c$ induces a parallax $\delta = g \cdot \Delta x_c / D$, resulting in a fringe phase change $\Delta \phi = 2\pi \Delta x_m / p_m$. Substituting the beat period $p_m$ and the displacement magnification relationship, and noting that the image-plane displacement of the grating component center is $\Delta u_{\text{trans}} = f \cdot \Delta x_c / D$, both expressions share the common factor $\Delta x_c / D$. Eliminating this factor yields $\Delta \phi = \frac{2\pi g}{p_r \cdot f} \cdot \Delta u_{\text{trans}}$. The linear scale factor between the phase change and the image-plane displacement depends solely on fixed parameters (interlayer gap $g$, rear period $p_r$, focal length $f$) and is entirely independent of the viewing distance $D$. More importantly, pure rotation does not contribute to parallax—rotation shifts both grating layers simultaneously without changing their relative position. Thus, only the translational component drives the phase change, and the above equation holds under general camera motion involving rotation (by subtracting the rotational component from the total displacement). Consequently, the verifier only needs to extract the cumulative phase signal and the pure translational displacement signal from the video and compute their Pearson correlation coefficient, without knowing any scene geometry parameters. The correlation coefficient naturally eliminates the scale factor, solely testing for the presence of linear coupling between the two signals. This principle of "detecting relationships rather than measuring values" is highly transferrable to other physical sensing domains.

2. FFT Phase Tracking: Extracting Sub-Pixel Fringe Displacements from Video Frames

Stably extracting the fringe phase from a single frame is a key engineering challenge for practical deployment. After perspective correction and CLAHE contrast enhancement, the fringe pattern is averaged along the fringe direction to yield a 1D intensity profile. Since the moiré beat frequency is the dominant periodic component in this profile, the corresponding frequency component of the 1D FFT directly encodes the lateral position of the fringes. The pipeline identifies and locks onto the dominant frequency bin $k^*$ in the first frame and directly reads the phase angle at this frequency in subsequent frames. To prevent $2\pi$ phase jumps, an incremental unwrapping strategy is adopted: the frame-to-frame phase difference is restricted to $[-\pi, \pi)$ via the wrap function and then accumulated, yielding a smooth cumulative phase signal $\Phi(t) = \sum \text{wrap}(\phi(\tau) - \phi(\tau-1))$. The advantage of this approach is that it is entirely parameter-free: it requires no manual tuning or template matching and exhibits natural robustness against contrast variations, partial occlusions, and slight blurring. Empirically, CLAHE contrast enhancement is a critical preprocessing step; in low-contrast or unevenly lit scenes, the FFT peak without enhancement is unstable and prone to hopping to noise components, causing phase tracking failure.

3. PnP Rotation Compensation: Recovering Pure Translational Signals from Mixed Camera Motions

In real video capture, camera motion is almost always a mixture of translation and rotation. Theoretical analysis indicates that only translation generates phase change through parallax, whereas pure rotation contributes nothing to fringe displacement. However, the raw pixel trajectory $\bar{c}(t)$ of the grating center is a superposition of both motions—directly calculating the correlation using this raw trajectory severely corrupts the signal. This work leverages the ArUco markers already present at the four corners of the grating component for rotation compensation. The four coplanar markers provide stable 3D-2D correspondences, allowing a PnP solver to recover the camera pose $(R_i, t_i)$ for each frame. By fixing the orientation of a reference frame to $R_{\text{ref}}$, the 3D center of the grating component is reprojected using each frame's camera position, yielding a displacement that contains only the translational component $\Delta u_{\text{trans}}$ when projected onto the direction perpendicular to the fringe lines. The elegance of this design lies in reusing the ArUco markers, which already serve the purpose of fringe region localization and perspective correction, to provide matching points for PnP—doubling the utility of a single marker without increasing hardware complexity. In evaluations, rotation compensation has a dramatic impact on the correlation coefficient: without compensation, the correlation of real videos plummets from 0.87 to approximately 0.5, making them nearly indistinguishable from generated videos, which proves this step is a backbone of the entire pipeline.

Key Experimental Results¶

Main Results¶

The paper collects 87 real videos (iPhone 15, 1080p@60fps) from 12 subjects across 29 scenes (25 indoor, 4 outdoor), covering three motion configurations: static camera/moving grating, moving camera/static grating, and simultaneous motion of both, with a shooting distance of approximately 1-3 meters. Meanwhile, 70 ideal verification videos are generated using the Blender Cycles physical rendering engine to establish the performance ceiling of the pipeline. On the AI generation front, the strongest attack setting is adopted—image-to-video (I2V) generation conditioned on the real first frame—filtering 92 manually verified "high-difficulty" negative samples from 317 candidate videos.

| Video Category | Quantity | Mean $|\rho|$ | Standard Deviation | Cohen's d vs. Ours (Real) | |---------|------|------|--------|------------------| | Real Videos (Indoor + Outdoor) | 87 | 0.87 | 0.14 | - | | AI-Generated Videos (All) | 92 | 0.57 | 0.21 | 1.71 | | Blender Physical Rendering | 70 | 0.99 | - | - | | Welch's t-test | $t(160) = 11.6, p < 10^{-20}$ | | | |

| Generative Model | Mean $|\rho|$ | |---------|---------| | Veo 3.1 | 0.58 | | Grok Imagine | 0.56 | | LTX-2 | 0.57 |

The performance across the three models is nearly identical, indicating that this is not an accidental failure of a specific model architecture but a systemic shortcoming of current video generation models in physical optics simulation. ROC analysis on the entire dataset yields an AUC of 0.8453, with an optimal decision threshold of $\tau^* \approx 0.927$. Note that this AUC is computed only on the 92 manually filtered high-difficulty negative samples. Out of the original 317 generated videos, 225 were excluded because the grating structures were severely deformed—these visually obvious failures do not pose a threat in practical deployment, hence the actual separation in real-world scenarios would be much higher.

Additional Robustness Tests¶

| Configuration | Windowed Mean $|\rho|$ | Range of Variation | Explanation | |------|--------|---------|------| | Orthogonal Bi-directional Fringes | 0.95 / 0.85 | [0.56, 0.99] | Tracking in both directions independently; both maintain high correlation | | In-plane Rotated Fringes | 0.76-0.81 | [0.03, 0.99] | Measurable after rotation compensation; window stability slightly decreases | | 45° Tilted Viewpoint | 0.72 | [0.14, 0.99] | Signal weakens under non-frontal viewpoints but does not drop to zero | | Dual-axis Fringes + Complex Rotation | 0.57-0.67 | [0.004, 0.99] | Most challenging configuration, but still adjudicated as real |

The results indicate that 3D tilt mainly reduces the signal-to-noise ratio without breaking the physical coupling relationship, and the windowed correlation remains much higher than that of the generated videos.

Key Findings¶

Rotation compensation is critical to system success: Removing rotation compensation causes the distribution of correlation coefficients for real and generated videos to overlap almost completely, destroying discriminability. PnP decomposition is a foundational component of the pipeline, rather than an optional refinement module.
Current state-of-the-art generative models fail across the board: Whether closed-source (Veo 3.1, Grok Imagine) or open-source (LTX-2), even in the highly advantageous I2V setup (with a real first frame input as a condition), the generated fringe motions fail the physical consistency check, and all three demonstrate virtually identical failures.
Zero pass rate in pure T2V generation: All three models completely fail to generate recognizable moiré pattern structures, indicating a fundamental lack of representation capacity for such fine periodic structures.
Splicing attacks are self-falsifying: Splicing a fringe region from a real video into another target video yields a transplanted phase signal that reflects the source video's camera-grating geometry. This fails to match the displacement trajectory of the target video, inevitably resulting in a low correlation coefficient.
Face-swapping attacks are a well-defined blind spot: If only the regions outside the grating component are modified (such as face-swapping), moiré authentication cannot detect it. However, this is a domain where existing face-swap detectors excel (82-97% accuracy), making the two approaches complementary parts of a multi-layered verification pipeline.

Highlights & Insights¶

"Physical invariants" are more durable than "finding statistical artifacts": The most core insight of this paper is to avoid entering an arms race against generative models (patching known artifacts) and instead let real physical processes generate deterministic signals that AI cannot forge. Even if future generative models learn to render moiré patterns, precisely simulating the frame-to-frame phase-displacement coupling still requires solving optical equations. This is equivalent to embedding a full-wave optical renderer into the generation pipeline—an extremely computationally expensive task that belongs to a different order of complexity compared to the "low-cost batch generation" attack scenario.
The elegance of calibration-free authentication: The moiré motion invariant leverages the fact that the unknown distance $D$ simultaneously appears in both the phase change and image displacement expressions and cancels out, making the Pearson correlation coefficient a distance-independent criterion. The verifier does not need to know the focal length, grating period, interlayer gap, or any other parameters—it only needs to detect whether a linear coupling between the two signals exists. This principle of "detecting relationships rather than measuring values" is highly transferrable to other physical sensing domains.
The double role of ArUco markers as clever engineering reuse: The ArUco markers at the four corners serve both fringe region localization/perspective correction and PnP pose estimation/rotation compensation. Achieving two crucial functionalities with a single set of markers minimizes hardware complexity. While a production system could replace these with more robust localization techniques, this reuse significantly lowered the engineering barrier at the prototype stage.
Categorized threat modeling provides clear boundaries: The paper systematically categorizes forgeries into T2V generation, I2V generation, splicing attacks, and face-swap edits, analyzing the security boundaries of each while explicitly acknowledging blind spots. Such honest and structured threat analysis is far more valuable than studies claiming to "solve all problems" and helps future researchers understand the precise positioning of this method within the forensic pipeline—it solves the question of "whether the video originates from a physical camera" rather than "whether any edits were made."

Limitations & Future Work¶

Dependence on relative motion: The authentication requires physical displacement between the camera and the grating to generate measurable phase shifts. When both are completely static, authentication cannot proceed. In practical designs, interactive prompts (e.g., "please sway the worn signature card slightly") must be introduced, which limits applications in static scenarios like fixed surveillance.
Vulnerability to future physical rendering engines: If attackers integrate full ray-tracing (e.g., using Blender Cycles or Unreal Engine to render gratings and synthesize them into backgrounds) into their generation pipelines, they could theoretically reproduce the correct coupling. However, such attacks require substantial manual engineering (scene matching, lighting alignment, viewpoint tracking), which stands in contrast to the "one-click mass forgery" threat model and is not currently an immediate threat.
Noise tolerance during deployment: Although the prototype performs exceptionally in controlled conditions, its performance under adverse conditions—such as outdoor environments, motion blur, low lighting, and H.264/H.265 compression on social media—requires large-scale verification. The authors propose multi-directional orthogonal grating patterns and window-level confidence filtering as future improvement directions.
Balancing user acceptance and scalability: Requiring users to wear or carry a grating component makes this an active, participatory verification method, which is less transparent than passive detection. It is acceptable in high-value scenarios like press conferences, video conferences, or judicial forensics (analogous to wearing a press badge), but large-scale adoption requires reducing component costs and improving aesthetic appeal.

vs. Passive Deep Learning Detectors (FaceForensics, CNN-based fingerprinting, etc.): Passive detectors rely on artifacts specific to generative models, leaving them disadvantaged in an arms race. The proposed physical signature does not rely on model-specific features, offering a more fundamental authentication mechanism. The two are complementary: passive detectors handle local edit (face-swapping) detection, while moiré signatures handle overall authentication of "whether the video originates from a physical optical system."
vs. Digital Watermarking (C2PA, SynthID): Watermarks can be stripped, re-encoded, or forged, and require embedding at the creation end. Moiré signatures are naturally generated at capture time without post-processing steps. However, watermarking scales to arbitrary content, whereas moiré signatures require specific hardware.
vs. Active Lighting Signatures (VeriLight, Michael et al.): Active methods require dedicated emitters to project structured light, increasing hardware complexity and power requirements. Conversely, the moiré signature is completely passive—relying solely on the geometric optical effects of double-layer gratings, which yields extremely low hardware costs (around $10 for lenticular film and printed paper) and requires no power source.
vs. Moiré Perception Studies (MoiréBoard, MoiréTracker, MoiréVision): These works utilize moiré patterns as measurement signals for pose estimation, deriving unknown camera poses from observations. This paper reverses this analytical direction, actively retaining and exploiting the eliminated distance term to construct an authentication invariant. The goal also shifts from "measurement" to "authentication."

Rating¶

Novelty: ⭐⭐⭐⭐⭐ Rediscovering moiré patterns—traditionally viewed as camera artifacts or pose estimation cues—as physical authentication signatures is highly original. The paradigm of "using physical laws rather than statistical detection for authentication" transcends the current passive detection framework.
Experimental Thoroughness: ⭐⭐⭐⭐ Systematically verifies 12 scenes, 3 motion configurations, and attacks from 3 different models, including strong I2V configurations and threat model analyses. However, the evaluation scale is relatively modest (87 + 92 videos) and lacks large-scale in-the-wild deployment data and quantitative testing under multi-platform compression or illumination degradation.
Writing Quality: ⭐⭐⭐⭐⭐ Rigorous theoretical derivation (from invariant derivation to engineering implementation and threat analysis, presenting a complete logical chain), exquisite illustrations, and an appendix providing a complete replication guide and prompt templates, making reproducibility excellent.
Value: ⭐⭐⭐⭐⭐ In an era of increasingly realistic AI-generated videos, this work provides a theoretically durable and practically deployable physical authentication solution. It holds immense potential for real-world scenarios, including video conferences, news verification, judicial forensics, and live broadcast authenticity assurance.