For years, digital forensics operated much like a standard quality-assurance pipeline. When an early Generative Adversarial Network (GAN) or a diffusion model synthesized an image or video, it left behind telltale signs: unnatural eye blinking, checkerboard patterns from upsampling layers, color mismatches along splicing boundaries, or high-frequency texture anomalies. Detectors were trained to hunt for these specific technical glitches, achieving high accuracy on standard benchmarks.
However, as noted by Isao Echizen et al. (2022), this approach trapped forensics in a perpetual, reactive cat-and-mouse game. Today, that game is reaching a breaking point. Generative models have evolved to achieve near-perfect physical and visual realism, suppressing the low-level spatial and temporal artifacts that early detectors relied upon. As synthetic content becomes visually indistinguishable from reality, traditional passive detectors are hitting a wall.
Why Artifact Hunting Fails
To understand why traditional detectors fail, we have to look at how modern generative models learn. Early networks lacked global coherence, leading to localized errors. But as blending, color correction, and anti-aliasing techniques matured, these surface-level glitches vanished.
When these artifacts are mitigated, the performance of traditional models collapses. For instance, Yuezun Li et al. (2020) demonstrated with the Celeb-DF dataset that when low-level splicing boundaries and color mismatches are mathematically cleaned up, the accuracy of state-of-the-art detectors drops from near-perfect to near-random levels. Furthermore, Anisha Pal et al. (2024) showed that forensic models heavily overfit to the low-level texture artifacts of specific generator versions, failing to generalize when faced with newer models or localized edits.
Even if a detector manages to catch a subtle high-frequency texture artifact in a pristine lab setting, these signals are incredibly fragile in the wild. As Xiaoyi Dong et al. (2022) point out, common real-world video processing operations—such as aggressive compression, resizing, or blurring during social media uploads—easily erase pixel-level traces. Relying on architectural flaws like upsampling errors is a losing strategy because, as argued by Renshuai Tao et al. (2025), those flaws disappear as generator architectures improve.
[Traditional Forensic Pipeline]
Raw Video ──> Look for Pixel Glitches (Blinking, Textures) ──> Fragile & Easily Erased by Compression
[Modern Invariant Pipeline]
Raw Video ──> Extract Underlying Invariants (3D Geometry, Biology) ──> Robust to Visual Smoothing
The Shift to Physical and Geometric Invariants
If pixels can be faked perfectly, what remains? The answer lies in the underlying physics and geometry of the real world. While a generative model can easily optimize for 2D pixel distributions that look realistic to a loss function, it struggles to maintain 3D physical consistency across frames.
Consider 3D head poses. Xin Yang et al. (2018) exposed deepfakes by calculating the geometric inconsistency between the estimated orientation of the central face (the synthesized region) and the entire head structure. Even if the pixels blend perfectly, the underlying 3D geometry is often slightly misaligned. Similarly, Andrea Ciamarra (2024) introduced an approach that estimates surface normals to uncover geometric "seams" that are invisible to the human eye but structurally impossible.
This principle extends to motion physics. The framework presented in MotionPhys (2026) tracks sparse optical-flow trajectories to measure physical properties like curvature and jerk. Because generative video models synthesize frame-by-frame or latent-by-latent without an explicit physics engine, they introduce subtle physical glitches in object trajectories that violate real-world mechanics, providing a durable signal for detection.
Exploiting Biological and Behavioral Signals
Another frontier bypasses technical artifacts entirely by focusing on human physiology and behavior—signals that generative AI cannot easily replicate because they require deep cross-modal or biological modeling.
One remarkable example is the visual heartbeat. Human skin undergoes subtle, periodic color changes due to blood flow, a phenomenon captured via remote photoplethysmography (rPPG). As shown by Hua Qi et al. (2020) and Jiahao Liang et al. (2022), while GANs and diffusion models can generate highly realistic skin textures, they fail to replicate the synchronized, rhythmic pulse of human blood flow across the face.
Behavioral biometrics offer a similar defense. Shruti Agarwal et al. (2023) developed a method that correlates person-specific facial movements with specific spoken words. A deepfake might generate a visually flawless mouth shape, but it often misses the idiosyncratic, semantic biometrics unique to how a specific individual moves their head and face when pronouncing certain words. Similarly, Trisha Mittal (2020) demonstrated that deepfakes frequently exhibit emotional mismatches between the audio track and visual expressions, exposing the lack of true affective coordination in synthetic media pipelines.
Proactive Verification and Deep Process Analysis
When passive detection of post-generation media becomes unreliable, the field shifts toward proactive defense or analyzing the mathematical nature of the generative process itself.
Proactive methods embed unforgeable signals before or during the synthesis process. For instance, Han Bao (2024) and Chang Liu (2024) propose watermarking frameworks that embed robust, invisible identifiers directly into model parameters or source media. These watermarks are engineered to survive generative manipulation, shifting the paradigm from catching flaws to verifying origin. On the cryptographic side, Mingyang Xie et al. (2022) introduce semantic signatures that sign the core meaning of an image (like identities and scene types) rather than its raw pixels, allowing benign edits while exposing deepfake alterations.
Alternatively, if one must analyze an unwatermarked image, detectors can look at how the model handles noise rather than looking at the final texture. Hengrui Lou et al. (2025) analyze the spatio-temporal distribution of how diffusion models fit noise during reconstruction. By focusing on these fundamental mathematical footprints rather than surface-level visual anomalies, forensics can achieve stable generalization even against entirely unseen generative models.